open-r1/OlympicCoder-7B

OlympicCoder-7B: A Specialized Code Generation Model

OlympicCoder-7B, developed by open-r1, is a 7 billion parameter model specifically fine-tuned for competitive programming. It builds upon the Qwen/Qwen2.5-Coder-7B-Instruct architecture and is designed to tackle complex coding challenges.

Key Capabilities

• Competitive Coding Performance: Achieves strong results on demanding benchmarks such as LiveCodeBench and the 2024 International Olympiad in Informatics (IOI'24).
• Fine-tuned on Decontaminated Data: Trained on a decontaminated version of the Codeforces dataset, enhancing its ability to solve algorithmic problems.
• C++ Post-training: The model was post-trained exclusively on C++ solutions, which may influence its performance on Python-centric benchmarks like LiveCodeBench, where it is considered partially out-of-domain.
• Chain-of-Thought (CoT) Prompting: The model's chat template is configured to encourage detailed chain-of-thought reasoning, prefilling the assistant's turn with a <think> token.

Good For

• Competitive Programmers: Ideal for generating solutions to challenging algorithmic problems encountered in contests like the IOI or platforms like Codeforces and LeetCode.
• Code Generation Tasks: Particularly effective for tasks requiring logical problem-solving and code implementation in a competitive context.
• Research in Code LLMs: Useful for researchers exploring model performance on advanced coding benchmarks and the impact of specialized training data.