A note of thanks · Specialist data

Savante and Aryabhata

Their mathematics specialist is a useful study of curated data, filtered traces, and verifiable rewards as one recipe. We have studied the published work, not reproduced its training result.

01 · The idea

What stayed with us

Curated domain data, filtered reasoning traces, supervised fine-tuning, model merging, and verifiable-reward reinforcement learning combine in a JEE mathematics specialist.

02 · The local translation

What we did with it

The study teaches how to separate initialization, data, filtering, training stages, reward, and evaluation before adapting a recipe to a Mac-sized target.

03 · The boundary

Where the comparison stops

The published result does not isolate each stage's contribution, and H100 training is not evidence of Mac-local reproducibility.