Summary
A Chinese forum post shares NVIDIA's recent paper on data engineering for improving large language model terminal (command-line) agent capabilities. The work introduces three core contributions: Terminal-Task-Gen, a lightweight pipeline for generating synthetic terminal tasks; Terminal-Corpus, a large-scale open-source dataset of terminal tasks; and Nemotron-Terminal, a family of open-source terminal agent models trained on top of Qwen3. The headline result is that the 32B model improves its score on Terminal-Bench 2.0 from 3.4% to 27.4%, achieving performance on par with much larger 480B-class models. The post includes links to the arXiv paper and the Hugging Face collection containing the open-source Nemotron-Terminal models, making it a useful pointer for researchers and practitioners interested in terminal agents, synthetic data generation, and efficient training of smaller models to match larger ones.
This forum post highlights a newly published NVIDIA paper on data engineering for scaling LLM terminal capabilities (i.e., agents that operate in the command-line/terminal environment).
Core Contributions
1. Terminal-Task-Gen: A lightweight synthetic task generation pipeline for terminal tasks.
2. Terminal-Corpus: A large-scale open-source dataset of terminal tasks.
3. Nemotron-Terminal: A family of open-source terminal agent models built on Qwen3.
Key Results
- The 32B model improves from 3.4% to 27.4% on Terminal-Bench 2.0, matching the performance of 480B-class large models.
Links
- Paper: https://arxiv.org/abs/2602.21193
- Open-source models: 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