Nemotron-Terminal-Corpus is a large-scale training dataset for terminal interaction capabilities released by NVIDIA in 2026. The associated paper is available at On Data Engineering for Scaling LLM Terminal Capabilities, and it aims to enhance the operational and reasoning performance of large language models in terminal environments through high-quality execution trajectories.

The dataset comprises approximately 366,000 high-quality execution trajectories, specifically designed for supervised fine-tuning (SFT), enabling models to perform practical tasks in complex domains such as data querying, model training, and code debugging. Models from the Nemotron-Terminal series, trained on this dataset, demonstrate excellent performance on the Terminal-Bench 2.0 benchmark. For example, Nemotron-Terminal-32B achieves 60.0% accuracy on data querying tasks, significantly outperforming the base model.

Dataset Composition

The dataset consists primarily of two parts, containing approximately 366k samples, all in the training (train) format:

• Dataset Adapters: High-quality mathematics, code, and software engineering (SWE) datasets converted into terminal format.

• Skill-based Synthetic Tasks: Novel tasks generated from a structured taxonomy of fundamental terminal skills, categorized into easy, medium, and mixed difficulty levels.

Citation

@misc{pi2026dataengineeringscalingllm,
      title={On Data Engineering for Scaling LLM Terminal Capabilities}, 
      author={Renjie Pi and Grace Lam and Mohammad Shoeybi and Pooya Jannaty and Bryan Catanzaro and Wei Ping},
      year={2026},
      eprint={2602.21193},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.21193}, 
}