Fine-Tuning Qwen3-27B for C-to-Rust Code Translation: A Three-Stage Curriculum of Pretraining, Debugging-Aware SFT, and Task-Specific SFT

A three-stage curriculum applied to Qwen3-27B improves C-to-Rust translation by combining Rust-centric continued pretraining with debugging-aware supervised fine-tuning. The approach leverages the microsoft/Verus_Training_Data dataset to teach the model to incorporate compiler and runtime feedback into the translation process, targeting improved idiomaticity and memory safety.