Starting from the same Nemotron base, fine-tuning produced two gold-medal-caliber tracks: one for solving problems, the other for writing proofs. The coding track used 22,000 problems and synthetic reasoning traces, then boosted scores through a generate–evaluate–revise loop, reaching 535.4/600 at the IOI; however, this was an unofficial, unproctored benchmark and did not count toward the official rankings.
The math track used over 410,000 examples covering proof generation, repair, and verification. In the end, it combined SFT and RL checkpoints because their strengths complemented each other, earning 30/42 at the IMO under official grading.
The math track used over 410,000 examples covering proof generation, repair, and verification. In the end, it combined SFT and RL checkpoints because their strengths complemented each other, earning 30/42 at the IMO under official grading.