Issue 23, Dec 28, 2025

The Curriculum Learning Trap

Staff Research Scientist interview at DeepSeek, and the interviewer asks:

โ€œWe have three massive datasets: ๐˜Ž๐˜ฆ๐˜ฏ๐˜ฆ๐˜ณ๐˜ข๐˜ญ ๐˜›๐˜ฆ๐˜น๐˜ต, ๐˜š๐˜ฐ๐˜ถ๐˜ณ๐˜ค๐˜ฆ ๐˜Š๐˜ฐ๐˜ฅ๐˜ฆ, and ๐˜ด๐˜ฑ๐˜ฆ๐˜ค๐˜ช๐˜ข๐˜ญ๐˜ช๐˜ป๐˜ฆ๐˜ฅ ๐˜”๐˜ข๐˜ต๐˜ฉ ๐˜ฑ๐˜ณ๐˜ฐ๐˜ฃ๐˜ญ๐˜ฆ๐˜ฎ๐˜ด. To build a State-of-the-Art Math reasoner, in what order do you feed this data during pre-training, and why?โ€

Donโ€™t say: โ€œJust shuffle them all together into one big dataset to avoid catastrophic forgetting.โ€

Why shuffling General, Code, and Math data together silently caps reasoning performance and how staged pretraining unlocks true chain-of-thought.

Share this trapShare on LinkedIn

The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.

Read it on Substack

Get the next one

Free on Substack. Unsubscribe in one click.