告别Docker：北大开源「迷你沙盒」，无容器也能训练SWE Agent

Core Viewpoint - The article discusses the development and advantages of SWE-MiniSandbox, a container-free framework designed to facilitate the training of Software Engineering Agents (SWE Agents) without the high costs associated with traditional container environments [3][4][6]. Group 1: Introduction to SWE-MiniSandbox - SWE-MiniSandbox aims to provide a low-cost framework for training SWE Agents, addressing the challenges posed by high resource requirements of container-based training environments [3][4]. - The framework eliminates the need for containers by utilizing process and file system isolation techniques, significantly reducing infrastructure and operational costs [6][7]. Group 2: Technical Innovations - The core optimization of SWE-MiniSandbox includes a combination of Chroot, Mount Namespaces, and Terminal Isolation, which collectively offer lower kernel overhead and faster performance compared to containers [9]. - A pre-caching pipeline is implemented to manage environment preparation efficiently, allowing for the reuse of compressed cache artifacts and reducing the environment size to approximately 5% of traditional container methods [7][21]. Group 3: Integration with Existing Tools - SWE-MiniSandbox integrates seamlessly with existing SWE tools such as SWE-Rex for terminal management, SWE-agent for task planning, and SkyRL for distributed reinforcement learning, enhancing its usability [17][18]. - The framework supports the creation of lightweight Python environments, significantly reducing the size of environment caches to around 100MB, compared to over 500MB for traditional conda environments [16][20]. Group 4: Experimental Results - Experimental results indicate that SWE-MiniSandbox achieves comparable training quality to traditional Docker frameworks while reducing environment preparation time to 25% of that required by Docker [21][22]. - In multi-node training scenarios, SWE-MiniSandbox maintains consistent environment startup speeds, demonstrating its scalability and efficiency [23]. Group 5: Future Prospects - Future developments for SWE-MiniSandbox may include the introduction of automated environment building processes, expansion to support more open-source SWE datasets, and optimization of training mechanisms to enhance efficiency [29].