How to reproduce a qwen-3b model that supports multiplication?

I have reproduced this project. The qwen-3b-instruct model can combine operations through addition and subtraction but fails to learn multiplication and division, even after I trained it for over 1,000 steps.The reward even get down by 0.1-0.2

From the author's WandB records, I found a multiply 3B model whose score approached 1 around 150 steps. How can I achieve this result? Is a model after SFT required?