Nemotron-Terminal-Corpus

# Terminal-Corpus: Large-Scale SFT Dataset for Terminal Agents Terminal-Corpus is a large-scale Supervised Fine-Tuning (SFT) dataset designed to scale the terminal interaction capabilities of Large Language Models (LLMs). Developed by NVIDIA, this dataset was built using the **Terminal-Task-Gen** pipeline, which combines dataset adaptation with synthetic task generation across diverse domains. ## 🚀 Key Results & Performance The high-quality trajectories in Terminal-Corpus enable models of various sizes to achieve performance that rivals or exceeds much larger frontier models on the **Terminal-Bench 2.0** benchmark. ### 1. Overall Performance Comparison Training on Terminal-Corpus yields substantial gains across the Qwen3 model family: | Model Size | Base Model (Qwen3) Accuracy | Nemotron-Terminal Accuracy | Improvement | | :--- | :---: | :---: | :---: | | **8B** | 2.5% ± 0.5 | **13.0% ± 2.2** | ~5.2x | | **14B** | 4.0% ± 1.3 | **20.2% ± 2.7** | ~5.0x | | **32B** | 3.4% ± 1.6 | **27.4% ± 2.4** | ~8.0x | The **Nemotron-Terminal-32B** (27.4%) outperforms the 480B-parameter **Qwen3-Coder** (23.9%) and **Gemini 2.5 Flash** (16.9%). **Nemotron-Terminal-14B** (20.2%) achieves higher accuracy than the 120B **GPT-OSS (high)** (18.7%). ### 2. Domain-Specific Breakthroughs The dataset unlocks functional utility in complex domains where base models previously showed near-zero capability: | Category | Qwen3-32B (Base) | Nemotron-Terminal-32B | | :--- | :---: | :---: | | **Data Querying** | 0.0% | **60.0%** | | **Model Training** | 0.0% | **50.0%** | | **Data Processing** | 5.0% | **50.0%** | | **Debugging** | 0.0% | **33.3%** | | **Software Engineering** | 5.0% | **31.7%** | ## 📂 Dataset Composition The released dataset contains approximately 366k high-quality execution trajectories split into two major streams: * **Dataset Adapters (~226k samples)**: Transformations of high-quality Math, Code, and Software Engineering (SWE) datasets into terminal-based formats. * **Skill-based Synthetic Tasks (~140k samples)**: Novel tasks generated from a structured taxonomy of primitive terminal skills. ## 📜 Citation If you use this dataset in your research, please cite the following work: ```bibtex @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}, }