rStar-Coder: How a 7-Billion-Parameter Model Mastered Competitive Programming Challenges

How a 7-Billion-Parameter Model Cracked Olympiad Programming: Inside Microsoft’s rStar-Coder unsplash.com/coding-laptop In May 2025, a research team quietly released a data set that changed the conversation around small language models (SLMs) and competitive programming. Named rStar-Coder, the project delivers 418 000 verified competition-grade code problems and 580 000 step-by-step reasoning solutions. When the team fine-tuned the modest Qwen2.5-Coder-7B on this data, the model leapt from 23 % to 62.5 % on LiveCodeBench—outperforming OpenAI o3-mini (low) and even QWQ-32B, a 32-billion-parameter powerhouse that generated the training rationales in the first place. This article explains—without marketing fluff—how the authors built the data …