OlympicCoder-7B
| Property | Value |
|---|---|
| Parameter Count | 7 Billion |
| Base Model | Qwen2.5-Coder-7B-Instruct |
| License | Apache-2.0 |
| Primary Language | English |
| GitHub Repository | https://github.com/huggingface/open-r1 |
What is OlympicCoder-7B?
OlympicCoder-7B is a specialized code generation model designed specifically for competitive programming tasks. It's a 7B parameter model fine-tuned on a carefully decontaminated version of the Codeforces dataset, with particular emphasis on handling complex algorithmic challenges similar to those found in programming olympiads.
Implementation Details
The model builds upon the Qwen2.5-Coder-7B-Instruct architecture and was trained using DeepSpeed Zero-3 optimization across 8 devices. The training process utilized careful hyperparameter tuning, including a learning rate of 4.0e-5 and a cosine learning rate scheduler with warmup.
- Batch size: 16 (2 per device with 8 gradient accumulation steps)
- Training duration: 10 epochs
- Optimization: Adam optimizer with betas=(0.9,0.999)
- Specialized chat template with
token implementation
Core Capabilities
- Strong performance on IOI'2024 benchmark problems
- Effective solution generation for LiveCodeBench challenges
- Primary focus on C++ programming solutions
- Integrated chain-of-thought reasoning through
token mechanism - Support for both competitive programming and algorithmic problem-solving
Frequently Asked Questions
Q: What makes this model unique?
OlympicCoder-7B stands out for its specialized focus on competitive programming challenges and its performance on prestigious benchmarks like the International Olympiad in Informatics. The model's unique chat template implementation with the
Q: What are the recommended use cases?
The model is particularly well-suited for competitive programming tasks, algorithmic problem-solving, and generating solutions for complex coding challenges. While primarily optimized for C++ solutions, it can handle Python code generation as well, though this might be considered partially out-of-domain.
