Hacker News new | past | comments | ask | show | jobs | submit
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.