Rufus-Air lays out an open, eight-stage recipe for post-training an LLM
The authors say they used public data and open-source components, without new human annotation or an in-house distillation teacher.
TLDR
The Rufus-Air paper describes an eight-stage post-training recipe for GLM-4.5-Air-Base. Its authors claim the resulting model improves on the official post-trained GLM-4.5-Air release and is competitive with similarly sized open models. An author says a small Amazon Search team built the recipe using public data and open-source components, rarely working with more than 400 H200s.
Combined views
3
1 Source, first seen ago