[2406.10305] Unlock the Correlation between Supervised Fine-Tuning and Reinforcement Learning in Training Code Large Language Models