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

Researchers introduced MindForge, a pipeline that turns open-source command-line programs into source-free training environments.

Using GLM-5.2, they generated program-synthesis data and fine-tuned Qwen3.6-27B.

The resulting model improved ProgramBench performance and also did better on seven other software engineering benchmarks.

The work points to a practical way to train smaller models for end-to-end software construction and coding-agent tasks.