This article reads like how to train a LLM

without a large corpus your pretrain is doomed to fail

Your post-train tricks hardly pays off if your base model doesn't scale.