open-r1/OlympicCoder-32B
Hugging Face
TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Mar 11, 2025License:apache-2.0Architecture:Transformer0.2K Open Weights Warm

OlympicCoder-32B by open-r1 is a 32 billion parameter code model fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct. It is specifically optimized for competitive coding benchmarks, demonstrating strong performance on challenging problems like those found in the International Olympiad in Informatics (IOI) and LiveCodeBench. The model excels at generating code solutions, primarily in C++, and is suitable for complex algorithmic tasks.

Loading preview...

OlympicCoder-32B: A Specialized Code Generation Model

OlympicCoder-32B, developed by open-r1, is a 32 billion parameter model fine-tuned from Qwen/Qwen2.5-Coder-32B-Instruct. Its primary focus is on competitive programming, where it has shown strong performance on demanding benchmarks. The model was post-trained on a decontaminated version of the Codeforces dataset, specifically utilizing C++ solutions generated by DeepSeek-R1.

Key Capabilities & Features

  • Competitive Coding Excellence: Achieves high performance on benchmarks like the 2024 International Olympiad in Informatics (IOI) and LiveCodeBench, which feature complex algorithmic problems.
  • C++ Optimization: While capable of generating Python, its training on C++ solutions means it is particularly adept at C++ code generation.
  • Chain-of-Thought (CoT) Prompting: The model's chat template is designed to encourage detailed thought processes, prefilling the assistant's turn with a <think> token to consistently enable long CoT reasoning.

Evaluation Highlights

Evaluated on two main competitive coding benchmarks:

  • IOI'24: Assessed on 6 highly challenging problems from the 2024 International Olympiad in Informatics.
  • LiveCodeBench: Tested on the v4_v5 subset of livecodebench/code_generation_lite, comprising 268 Python problems. It's important to note that performance on LiveCodeBench is partially out-of-domain as the model was primarily trained on C++ solutions.

Good For

  • Developers and researchers focused on competitive programming.
  • Generating solutions for complex algorithmic problems, especially in C++.
  • Exploring advanced code generation with explicit chain-of-thought reasoning.
Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p