MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis

Benchmark Model Rank Results
code-generation-on-deepsweMindForge-27B (Mini-SWE-Agent)–Pass@1: 0.1592
code-generation-on-programbenchMindForge-27B (Mini-SWE-Agent)–Average Test Pass Rate: 49.51
code-generation-on-swe-bench-multilingualMindForge-27B (Mini-SWE-Agent)–Accuracy: 67.77
code-generation-on-swe-bench-proMindForge-27B–Resolved: 51.34
code-generation-on-swe-bench-verifiedMindForge-27B (Mini-SWE-Agent)–Accuracy: 73.84