MindForge: Significantly Boosting Small Models' From-Scratch Coding Abilities

centre-for-swe · hf · 2026-07-31

While coding agents have made progress in modifying existing codebases, building complete programs from scratch remains a major challenge, with frontier models fully resolving less than 1% of ProgramBench tasks.

To address this, researchers introduced MindForge, a training pipeline that converts open-source command-line programs into source-free environments exposing only compiled executables and documentation.

Experiments show that fine-tuning a Qwen3.6-27B model on synthesis trajectories generated by a GLM-5.2 teacher agent increases its ProgramBench pass rate from 37.98% to 49.51%, rivaling larger frontier models. Furthermore, the fine-tuned model consistently improves across seven unseen software engineering benchmarks, including cross-language issue resolution, bug fixing, and repository generation.

Original post →

More from coding & agent

coding & agent channel →