The project trains specialized LLMs on binary-source code pairs to perform decompilation, enabling recovery of human-readable code from compiled executables. It provides trained models (e.g., SK-Decompile for two-phase binary decompilation), evaluation benchmarks (decompile-bench with 70K function pairs), and datasets for training. The work focuses on improving decompilation accuracy through model fine-tuning and structured approaches combining structure recovery with identifier naming.