This post was originally published on Substack. Click the link to read the full article.
Why shuffling General, Code, and Math data together silently caps reasoning performance and how staged pretraining unlocks true chain-of-thought.
Read the full article on Substack