SWE Agent Data Collection and Execution Environment

11/24/2025

Introduces a data curation pipeline and repository sandbox generator for SWE agent training. Details methodology for sourcing repositories from PyPI, GitHub stars, and multimodal codebases, applying strict quality and activity thresholds. Describes the architecture of the code execution environment, which automates differential testing of Python repositories within isolated Docker environments, including dependency resolution, Dockerfile generation, LLM-based build failure fixes, test discovery, execution, and classification of test outcomes (FAIL_TO_PASS, PASS_TO_PASS).

11/24/2025

The post describes refactoring the OpenHands evaluation workflow to dynamically load arbitrary Docker images and parse metadata at runtime, enabling support for thousands of diverse Python repositories. It also introduces a unified rollout architecture that maintains a persistent runtime context, eliminating container duplication and reducing overhead by approximately 95% for large-scale RL training. This enables multi-turn interaction support, fast rollout generation, sparse reward evaluation, and comprehensive trajectory logging.