A naukri.com initiative
Nvidia
4w
235
Image Credit: Nvidia
Math Test? No Problems: NVIDIA Team Scores Kaggle Win With Reasoning Model
- The NVIDIA team recently won the AI Mathematical Olympiad's competition on Kaggle with an AI reasoning model.
- Team members from different locations participated in the competition, continuously submitting and improving their model.
- Their model excelled at generalizing to unseen data, correctly answering 34 out of 50 Olympiad questions within five hours.
- Named NemoSkills, the team utilized NeMo-Skills for LLM training and used a combination of natural language reasoning and Python code execution.
- The team's winning model, based on Qwen2.5-14B-Base, underwent knowledge distillation to optimize performance.
- They employed techniques like asynchronous processing and FP8 quantization using TensorRT-LLM for faster inference.
- The model performed exceptionally well on the competition's final dataset, showcasing its generalizability and avoiding overfitting.
- The team plans to release a technical report, share their techniques, and contribute to NeMo-Skills pipelines on GitHub.
- Using tools across the NVIDIA software stack, the team's optimizations have been integrated into open-source libraries.
- Team members regained titles such as Kaggle World Champion, Kaggle Grandmaster, and directed the prize towards charitable organizations.
- The team includes Igor Gitman, Darragh Hanley, Christof Henkel, Ivan Moshkov, Benedikt Schifferer, Ivan Sorokin, and Shubham Toshniwal.
Read Full Article
14 Likes
For uninterrupted reading, download the app