Nemotron-RL-Agentic-Terminal-Pivot-v1

## Dataset Description The **Nemotron-RL-Agentic-Terminal-Pivot-v1** dataset provides training samples for reinforcement learning of command-line ("terminal use") LLM agents with the **terminus_judge** environment in **NeMo Gym**. Each record is a single agent decision point extracted from a successful agent trajectory on a terminal task: - `responses_create_params.input` — the prompt: the task instruction plus the terminal interaction history (prior agent actions and terminal outputs) up to the decision point; - `expected_answer` — the reference next action taken by the teacher agent at that point (a JSON object in Terminus-2 format with `analysis`, `plan`, `commands[].keystrokes`, and `task_complete` fields); - `agent_ref` — routing to the NeMo Gym `terminus_judge` resources server, which scores a policy model's generated action against the reference action to produce a reward for Reinforcement Learning from Verifiable Reward (RLVR). The samples are drawn from **ATCB**, a collection of terminal-bench-style tasks; the dataset covers 630 unique ATCB seed tasks. Each seed task is a containerized Linux environment with a natural-language instruction, a hidden reference solution, and an automated verifier. The tasks emphasize realistic, long-horizon operational work in the terminal: building, repairing, and reconciling data pipelines; auditing and recovering from injected faults in logs, ledgers, caches, and telemetry stores; diagnosing crashed services, schedulers, and CI builds; and security/compliance workflows (access-log audits, SIEM triage, PII handling, TLS and authentication fixes) — set in scenarios spanning industrial and embedded systems (SCADA, firmware/OTA, cold-chain, satellite), HPC clusters, healthcare, finance, and media archives. Trajectories were generated by running the seed tasks with the **Terminus-2** agent (v2.0.0) in the Harbor execution harness, with **GLM-5.1** (`zai-org/GLM-5.1`) as the teacher model. Only trajectories that passed the task's automated verifier were kept (at most 5 successful trajectories per task); each valid assistant turn of a kept trajectory was then converted into one next-action training sample. Degenerate turns (over-long command strings, duplicate prompts, post-completion confirmations) and records whose observation text captured collection-infrastructure details were filtered out. All 630 seed tasks are ATCB tasks, and every released sample derives exclusively from ATCB trajectories. Internal collection metadata (provenance identifiers, pipeline and infrastructure fields) was likewise removed from every record; each record retains only the minimal fields documented under Dataset Quantification. This dataset was utilized in the development of the NVIDIA Nemotron family of models: the same 630-task ATCB dataset was used for the reinforcement-learning post-training of both **NVIDIA Nemotron Ultra** and **NVIDIA Nemotron 3.5 Lightning**. This dataset is ready for commercial or non-commercial uses. ## Dataset Owner(s) NVIDIA Corporation ## Dataset Creation Date Trajectories collected and samples constructed: 05/2026. Release preparation: 07/21/2026. ## Versioning v1.0 Previous Version(s): no previous version. Relationship to Previous Version(s): Not applicable. ## License/Terms of Use This dataset is licensed under Creative Commons Attribution 4.0 International (CC-BY-4.0). ## Intended Usage Post-training large language models for agentic terminal use with reinforcement learning. The dataset is designed to be consumed by the NeMo Gym `terminus_judge` resources server (https://github.com/NVIDIA-NeMo/Gym/tree/main/resources_servers/terminus_judge): during training, a policy model is prompted with `responses_create_params.input`, and the judge scores the generated next action against `expected_answer` to produce a reward. The samples can also be used for supervised fine-tuning of terminal agents or for offline analysis of terminal-agent behavior. ## Dataset Characterization **Data Collection Method** * Hybrid: Manually-Collected, Synthetic, Automated — seed tasks (instructions, environments, verifiers) in the ATCB collection were authored by a data vendor (Human); agent trajectories were generated by the GLM-5.1 teacher model driving the Terminus-2 agent in containerized task environments (Synthetic); execution, verification, and per-turn sample extraction were performed by automated pipelines (Automated). **Labeling Method** * Hybrid: Synthetic, Automated — the reference next action in each sample is teacher-model output (Synthetic); trajectory-level success labels come from each task's automated verifier script, and only verifier-passing trajectories were retained (Automated). No human labeling and no filtering model were used. ## Dataset Format Modality: Text. Format: JSON Lines (`.jsonl`), one training sample per line, UTF-8. Records are compatible with the NeMo Gym `terminus_judge` environment (https://github.com/NVIDIA-NeMo/Gym). ## Dataset Quantification Record count: 31,111 training samples (630 unique seed tasks; 2,716 distinct source trajectories; median 45 samples per task). Feature count: 8 top-level fields — 7 data fields (including `responses_create_params`, `expected_answer`, `agent_ref`) plus a `metadata` object with 5 sub-fields. Average prompt length ≈ 39,900 characters; average reference action ≈ 970 characters. Total data storage: ~1.4 GB (JSONL, uncompressed). ## Reference(s) - NeMo Gym: https://github.com/NVIDIA-NeMo/Gym - terminus_judge resources server: https://github.com/NVIDIA-NeMo/Gym/tree/main/resources_servers/terminus_judge - Terminal-Bench: https://github.com/laude-institute/terminal-bench - Terminus-2 agent / Harbor harness: https://github.com/laude-institute/harbor - GLM-5.1: https://huggingface.co/zai-org/GLM-5.1 ## Ethical Considerations NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse. Task environments are synthetic, containerized scenarios authored by a data vendor; personal-appearing data in task content (names, e-mail addresses, credentials) was authored as fictional scenario material. Released text was additionally scanned for real personal and infrastructure information: no real personal data was identified, internal-only collection metadata (provenance identifiers, pipeline and infrastructure fields) was removed from all records, and records whose observation text captured collection-infrastructure details were dropped prior to release. Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).