Rufus-Air Post-Training Recipe for 106B GLM-4.5-Air-Base

Rufus-Air provides a reproducible eight-stage post-training pipeline for the 106B GLM-4.5-Air-Base model, covering SFT, Reasoning RL, Coding RL, and Agentic RL. The recipe emphasizes difficulty filtering and reward reliability to move from basic instruction-following to advanced agentic capabilities without new human annotations.