Terminal-Corpus
收藏资源简介:
Terminal-Corpus是由英伟达团队构建的大规模开源终端任务数据集,旨在提升大语言模型在命令行环境中的交互能力。该数据集通过轻量级合成任务生成管道Terminal-Task-Gen构建,包含数据集适配和合成任务生成两阶段数据,覆盖数学推理、代码生成和软件工程修复等领域。其核心创新在于将现有基准转化为终端提示,并结合技能分类法生成针对性场景,最终通过Docker化环境交互产生解决方案轨迹。该数据集主要用于训练Nemotron-Terminal模型家族,显著提升了终端任务基准测试Terminal-Bench 2.0的性能表现,解决了终端代理训练数据稀缺和生成复杂度高的核心瓶颈。
Terminal-Corpus is a large-scale open-source terminal task dataset developed by the NVIDIA team, designed to improve the command-line interaction capabilities of large language models (LLMs). This dataset is built via the lightweight synthetic task generation pipeline Terminal-Task-Gen, which consists of two stages: dataset adaptation and synthetic task generation, covering domains such as mathematical reasoning, code generation, and software engineering repair. Its core innovation lies in converting existing benchmarks into terminal prompts, generating targeted scenarios in conjunction with a skill taxonomy, and ultimately generating solution trajectories through interactions within Dockerized environments. This dataset is primarily utilized for training the Nemotron-Terminal model family, significantly boosting the performance of the terminal task benchmark Terminal-Bench 2.0, and resolving the core bottlenecks of scarce training data and high generation complexity for terminal AI agents.



