Loading post…
Hugging FaceAleksander, Igor Gitman, Sean Narenthiran, Mehrzad Samadi, Somshubra MajumdarAnnouncement
One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO
Summary
Starting from Nemotron 3, the team applied supervised fine-tuning (SFT) and reinforcement learning (RL) to create specialists for competitive programming and olympiad mathematics. At the IOI, the Nemotron-3-Ultra-CC model, trained with SFT on 22,000 curated problems and paired with GenCorrect iterative refinement, scored 535.4 of 600 points, exceeding both the gold threshold and the top human score. At the IMO, a combination of SFT and RL checkpoints scored 30 of 42 points by generating, critiquing, and refining natural-language proofs without external tools. Both projects followed a shared recipe: start with a strong base, curate domain-specific data, apply standard post-training, and tie the model to an inference loop that improves candidates. The results show that specialization and test-time compute together produce gold-level performance, and all models, data, and inference pipelines are available on Hugging Face.
Context
The IOI and IMO are notoriously difficult competitions; the IOI tests algorithmic coding under tight constraints, while the IMO tests rigorous mathematical proof writing. Success at both from a single model family demonstrates broad domain adaptability.
Approach / What changed
The approach starts with a strong Nemotron 3 base model, curates domain-specific problems and reasoning traces, applies supervised fine-tuning (SFT) and reinforcement learning (RL), and pairs the specialist model with a test-time inference loop that generates, evaluates, and refines candidates. For the IOI, the inference strategy is GenCorrect (generate-evaluate-refine). For the IMO, the system uses multiple checkpoints (general, SFT, RL) in a generate-verify-refine pipeline without external tools.
Takeaways
Nemotron 3 can be fine-tuned to gold-medal level at both IOI and IMO using standard SFT and RL, without building a new foundation model.
For the IOI, the GenCorrect iterative generate-evaluate-refine strategy was key; it lifted the Nano model from 291 to 468 points after RL, crossing the gold threshold.
For the IMO, using complementary SFT and RL checkpoints together was more valuable than sampling more from one checkpoint; the final system scoring 30/42 used all three.
FAQ
What is GenCorrect and how did it help the IOI model?
GenCorrect is an iterative generate-evaluate-refine strategy: the model proposes solutions, evaluates them, and refines the best ones over multiple rounds. It helped the Nano model cross the gold threshold from 291 to 468 points after RL.
How did the IMO system combine multiple checkpoints?
The system used three checkpoints: a general Nemotron 3 Ultra, an SFT specialist trained on 414,890 examples, and an RL specialist trained on 9,597 frontier problems. They generated candidates, scored, critiqued, and refined; the final submission was selected in a high-compute stage.
Are the Nemotron IOI and IMO models publicly available?
Yes. The IOI model Nemotron-3-Ultra-CC, the IMO SFT and RL checkpoints, training datasets, Nemotron-IMO-Bench, and the NeMo-Skills inference pipelines are all available on Hugging Face.
Source and freshness
- Source
- huggingface.co
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.