aws-prototyping/codefu-7b-v0.1
CodeFu-7B-v0.1 is a 7.61 billion parameter model developed by AWS WWSO Prototyping, built on the DeepSeek-R1-Distill-Qwen-7B base. It is specifically trained using Reinforcement Learning for competitive programming tasks, excelling at algorithmic reasoning and generating efficient C++ solutions. The model achieves 13.7% Pass@1 on the USACO benchmark, outperforming models more than four times its size and showing a >10x improvement over its base model.
Loading preview...
CodeFu-7B-v0.1: Competitive Programming Specialist
CodeFu-7B-v0.1 is a 7.61 billion parameter model from AWS WWSO Prototyping, engineered for competitive programming and algorithmic problem-solving. Built upon the DeepSeek-R1-Distill-Qwen-7B base, this model leverages Reinforcement Learning (RL) to develop strong algorithmic reasoning capabilities and generate effective C++ solutions.
Key Capabilities & Differentiators
- Exceptional Performance for its Size: Achieves 13.7% Pass@1 on the challenging USACO benchmark, a leading score among 7B models.
- Outperforms Larger Models: Demonstrates superior performance compared to its 32B parameter base model (13.7% vs 11.7% Pass@1) and other 32B models like DeepSeek-R1-Distill-Qwen-32B.
- Significant Improvement Over Base: Shows a remarkable >10x performance increase over its 7B base model (13.7% vs 1% Pass@1), highlighting the effectiveness of its RL training approach.
- RL Training on Problem Statements: Uniquely trained solely on competitive programming problem statements from the DeepMind CodeContest dataset, without access to ground-truth solutions.
- Structured Prompting: Optimized for prompts that request both thought processes (within
<think>tags) and C++ code solutions (within<code>tags).
Good For
- Algorithmic Problem Solving: Ideal for tasks requiring complex algorithmic reasoning and efficient code generation in C++.
- Competitive Programming: Suited for generating solutions to problems found in contests like the USA Computing Olympiad.
- Code Generation with Reasoning: Useful in scenarios where both the thought process and the final code solution are desired.