Cerebras

This post details the methodology for constructing a high-quality, diverse corpus of real-world software projects for training SWE agents. It describes sourcing repositories from three pipelines: PyPI-linked projects (107,000 instances), high-impact GitHub repositories (50,000 instances), and multimodal codebases (3,500 instances). The post also outlines the architecture of the code execution environment, which automates end-to-end differential testing of Python repositories by setting up isolated workspaces, analyzing repository structure and dependencies, generating optimized Dockerfiles, building environments with LLM-based fixes for build failures, performing robust test discovery with iterative auto-correction, executing baseline and patched tests in isolated Docker runs, and comparing results to classify tests as FAIL_TO_PASS or PASS_TO_PASS. A reversal phase rechecks classifications by undoing the main patch, ensuring accuracy. The system produces comprehensive artifacts including logs, Docker files, summaries, and detailed JSON results.