Summary
NVIDIA has published a new paper on data engineering for improving LLM terminal (command-line) agent capabilities. The work introduces three main contributions: Terminal-Task-Gen, a lightweight synthetic task generation pipeline; Terminal-Corpus, a large-scale open-source terminal task dataset; and Nemotron-Terminal, a family of open-source terminal agent models trained on the Qwen3 base. Key results show the 32B model improving from 3.4% to 27.4% on Terminal-Bench 2.0, matching the performance of much larger 480B models. The paper, models, and datasets are publicly available: the paper is on arXiv (https://arxiv.org/abs/2602.21193) and the Nemotron-Terminal model collection is hosted on Hugging Face (https://huggingface.co/collections/nvidia/nemotron-terminal). This demonstrates that targeted synthetic data engineering for terminal environments can dramatically close the gap between mid-sized open models and frontier-scale systems.
NVIDIA recently published a paper on data engineering for terminal (command-line) agent capabilities in LLMs.
Core Contributions
1. Terminal-Task-Gen: A lightweight synthetic task generation pipeline
2. Terminal-Corpus: A large-scale open-source terminal task dataset
3. Nemotron-Terminal: A family of open-source terminal agent models trained on the Qwen3 base
Key Results
The 32B model improved from 3.4% to 27.4% on Terminal-Bench 2.0, matching the performance of a 480B-scale model.
Links
- Paper: https://arxiv.org/abs/2602.21193
- Open source: https://huggingface.co/collections/nvidia/nemotron-terminal
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/177168806