Rufus-Air Turns LLM Post-Training into a Reproducible Pipeline

Introduction

The race to improve language models is increasingly shifting from pre-training scale to the quality of post-training. Rufus-Air is notable because it presents post-training as an open and reproducible engineering pipeline rather than a collection of isolated tricks. Built on GLM-4.5-Air-Base (106B-A12B), the recipe explains how a base model can be progressively developed for reasoning, coding, instruction following, and agentic tasks through coordinated data, rewards, and stage ordering.

Eight stages with a deliberate progression

Rufus-Air uses a serial pipeline:

• SFT establishes a broad capability floor with diverse, high-quality examples.
• Reasoning RL improves reasoning behavior.
• Coding RL targets programming tasks with reinforcement learning.
• Instruction-Following RL strengthens adherence to user requests.
• General Agent expands the model toward general agent tasks.
• Coding Agent focuses on agentic coding scenarios.
• Search Agent develops search-oriented agent behavior.
• RLHF introduces a final preference-oriented reinforcement learning stage.

The order is central to the recipe. Tasks with clear outcomes and dependable verification are handled earlier, while more complex tasks that require broader judgment come later. In parallel, the training moves from hard, relatively reliable rewards toward softer signals produced by judges. This progression is intended to reduce instability caused by ambiguous optimization targets.

Four lessons beyond the model name

First, SFT is more than a warm-up. The authors argue that diverse, high-quality supervised data defines the capability floor from which later reinforcement learning can improve.

Second, prompt difficulty needs to be controlled. Easy prompts provide little useful learning signal, while excessively difficult ones may push training outside a productive range. Filtering for tasks that are challenging but learnable becomes a practical lever for RL efficiency.

Third, reward reliability offers a useful rule for deciding stage order. Executable tests and other explicit checks are suitable for earlier phases; judge-based rewards are better introduced after the model has developed stronger basic abilities.

Fourth, infrastructure is part of the recipe. Data pipelines, training systems, evaluation procedures, and transitions between stages all affect the outcome. Reproducibility therefore cannot be guaranteed by naming an algorithm alone.

Why it matters

Rufus-Air suggests that open post-training depends not only on a capable base model, but also on organizing data curation, reward design, and engineering execution into a stable process. The authors report improvements over the official post-trained GLM-4.5-Air release and competitiveness with similarly sized open models. They also state that the project relies on open-source components and public data, without new human annotation or an in-house distillation teacher.

For researchers, the staged design creates a framework in which each capability transition can be examined separately. For model builders, the broader lesson is straightforward: post-training is not simply a matter of adding RL. It is a systems process built around capability foundations, reward trustworthiness, and task complexity.

Source: Hugging Face Daily Papers