MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis
MindForge turns existing CLI programs into source-hidden training tasks, then uses teacher-generated synthesis traces to improve a 27B coding model.
The pipeline exposes only a compiled reference executable and documentation, aiming at from-scratch program construction rather than edits to an existing repo. Fine-tuning Qwen3.6-27B on GLM-5.2 trajectories raised its ProgramBench average test pass rate from 37.98% to 49.51%. The paper also reports gains over the base model across seven unseen software engineering benchmarks, including repository generation, translation, bug fixing, feature work, and multilingual issue resolution. HF Daily Papers' note
The pipeline exposes only a compiled reference executable and documentation, aiming at from-scratch program construction rather than edits to an existing repo. Fine-tuning Qwen3.6-27B on GLM-5.2 trajectories raised its ProgramBench average test pass rate from 37.98% to 49.51%. The paper also reports gains over the base model across seven unseen software engineering benchmarks, including repository generation, translation, bug fixing, feature work, and multilingual issue resolution. HF Daily Papers' note
score 5