Most terminal-agent datasets are either short-form command mappings or synthetic single-step examples; this dataset captures long-horizon, verifier-passing agent trajectories and exposes each assistant decision as a training sample, making it directly usable for RL from verifiable reward and for supervised fine-tuning of terminal agents.
What Sets It Apart
- Per-decision samples extracted from successful multi-step trajectories: each record contains the full natural-language task instruction plus the full terminal interaction history up to the decision point, and an expected next action encoded as a Terminus-2 JSON object (analysis, plan, commands[], task_complete). This preserves long-context, multi-turn decision structure rather than isolated input–output pairs.
- Realistic, verifier-validated tasks: samples derive exclusively from 630 ATCB seed tasks (containerized Linux scenarios) emphasizing operational, safety, and long-horizon workflows (debugging, recovery, security audits, data-pipeline fixes). Only trajectories that passed each task's automated verifier were retained.
- RL-ready judge integration: records are shaped to work with NeMo Gym's terminus_judge resources server — during post-training RL, a policy's generated action is scored against the teacher reference to produce verifiable rewards, avoiding subjective human reward labels.
- Traceable collection pipeline and teacher model: trajectories were generated by the Terminus-2 agent using GLM-5.1 as the teacher; degenerate turns and internal provenance metadata were filtered out to produce compact, usable samples.
Key dataset facts
- Record count: 31,111 training samples; 630 unique seed tasks; 2,716 distinct source trajectories; median 45 samples per task.
- Average prompt length ≈ 39,900 characters; average reference action ≈ 970 characters; total uncompressed storage ≈ 1.4 GB. Format: JSONL compatible with NeMo Gym terminus_judge.
- License: Creative Commons Attribution 4.0 (CC-BY-4.0). Trajectories collected 05/2026; release prepared 2026-07-21.
Who it's for and trade-offs
Great fit if you need to fine-tune or RL-post-train an LLM to perform multi-step terminal workflows under an automated verifier, or if you need offline analysis of agent decision behavior across long contexts. It is also directly consumable by NeMo Gym/terminus_judge workflows. Look elsewhere if you need human-labeled intent classifications, large-scale diverse natural dialogs, or lightweight single-command datasets: average prompt sizes and verifier integration make training resource-intensive and require infrastructure (terminus_judge) to realize RL-from-verifiable-reward. The dataset reflects teacher-model behaviors (GLM-5.1) rather than human annotations, which is efficient but inherits teacher biases and failure modes.
