> ## Content Index > Fetch the complete content index at: https://www.stirup.si/llms.txt > Use this file to discover other available public pages before exploring further. # Model Adaptation for IMO and IOI Competitions - URL: https://www.stirup.si/model-adaptation-for-imo-and-ioi-competitions-67f5f616/ - Published: 2026-10-13T15:04:52.000Z - Updated: 2026-10-13T15:04:52.000Z - Description: Hugging Face refined Nemotron 3 using SFT, RL, and synthetic reasoning traces to achieve gold-level results in IMO 2026 and IOI 2026 competitions. - Author: Satish Das - Tags: AI, News, Stirup Research ## Model Adaptation for Competitions Model adaptation for competitions involved a structured approach to refine Nemotron 3 into systems capable of achieving gold-level results at both IMO 2026 and IOI 2026\. According to [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si), the process utilized supervised fine-tuning (SFT), reinforcement learning (RL), and feedback-driven inference. [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si) outlines a four-part strategy, beginning with a strong Nemotron base model. For competitive programming, [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si) highlights the curation of 22,000 problems and the generation of synthetic reasoning traces to train two specialists, focusing on general coding ability and advancing to IOI gold. ## Model Specifications and Training Methods The Nemotron-3 series includes two distinct models with varying parameter counts and training approaches. Nemotron-3-Nano-CC features 30 billion total parameters, of which 3 billion are active, and underwent both supervised fine-tuning (SFT) and reinforcement learning (RL) during training. In contrast, Nemotron-3-Ultra-CC has a significantly larger scale, with 550 billion total parameters and 55 billion active parameters, and was trained using only SFT. According to [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si), starting from Nemotron 3 Ultra, the team trained one specialist using SFT and another using RL, highlighting a divergence in training strategies for different applications. These specifications suggest that the Nano variant is optimized for efficiency and adaptability, while the Ultra variant is designed for broader, more complex tasks. The training methods applied to each model reflect their intended use cases and performance goals. ## Performance Outcomes and Thresholds In the IOI 2026 competition, Nemotron-3-Ultra-CC with SFT and GenCorrect achieved a score of 535.4/600, surpassing the gold threshold of 361.12 and the top human score of 498.27\. This result was obtained under the same time, internet-access, and submission constraints as human contestants. For the IMO 2026, Nemotron 3 Ultra general, SFT, and RL checkpoints in a generate-verify-refine system scored 30/42, exceeding the official gold threshold of 29\. The model earned full credit on four of the six problems. With GenCorrect, the iterative generate-evaluate-refine strategy reached 468 points, crossing the gold threshold of 438.3\. These outcomes highlight the effectiveness of specialized training and iterative refinement techniques in enhancing model performance on high-stakes competitions. ## Training and Inference Strategies Fine-tuning and test-time compute work together to enhance model performance, as demonstrated in the IOI 2025 Hugging Face post. At IOI, GenCorrect leveraged fine-tuning to turn initial gains into larger improvements over multiple feedback rounds. The medals achieved were not the result of fine-tuning alone, nor were they the result of brute-force sampling alone. This highlights the importance of combining strategic fine-tuning with effective test-time compute to maximize results. These findings underscore the need for a balanced approach in training and inference strategies, where neither component is relied upon exclusively. The success at IOI and IMO illustrates that the most effective outcomes arise from the synergy between these two methods. ## Resource Availability and Benchmarking Resource availability and benchmarking play a critical role in evaluating the performance of fine-tuned models like Nemotron. The Nemotron Labs IMO 2026 collection includes SFT and RL checkpoints, training datasets, and Nemotron-IMO-Bench, a benchmark with 200 olympiad-level problems [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si). For competitive programming, the Nemotron-3-Ultra-CC model is accessible on Hugging Face, with the IOI paper detailing the training recipe and GenCorrect methodology [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si). These resources collectively offer rigorous testing grounds for Nemotron, demonstrating its potential to be fine-tuned into domain specialists capable of solving problems at the frontier of human competition [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si). ## Sources and reporting notes - [Hugging Face](https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026?ref=stirup.si) Based on published source material. No independent hands-on testing is claimed.