September 24, 2026
A local model can learn a domain: a 4-stage experiment with Qwen 3.5 4B
On Sep 24, the experiment's author described four stages of training local Qwen 3.5 4B on a new domain through CPT and RAG. They compared knowledge embedded in the model with RAG, then combined both approaches.
On Sep 24, the experiment's author described four stages of training local Qwen 3.5 4B on a new domain through CPT and RAG. They compared knowledge embedded in the model with RAG, then combined both approaches.
In the final part, the author describes an eval approach in which the fine-tuned model responds according to a strict schema rather than with English phrases.
