Confucius3-Math: Specialized K-12 Math Reasoning LLM
Confucius3-Math is a 14-billion parameter large language model developed by the NetEase Youdao AI Team, uniquely focused on K-12 mathematics education. It distinguishes itself from general-purpose LLMs through an RL-only post-training process, incorporating a novel data scheduling policy and an improved group-relative advantage estimator.
Key Capabilities and Differentiators
- SOTA Performance on Math Tasks: Achieves state-of-the-art results on Chinese K-12 math problems, surpassing larger models on benchmarks like CK12-MATH (96.24%), GAOKAO-Bench (98.46%), CMATH (96.13%), MATH-500 (98.80%), and AIME 2024 (81.15%).
- Cultural & Curriculum Alignment: Specifically optimized for China's national mathematics standards and problem-solving methodologies, ensuring relevance and accuracy for its target demographic.
- Cost-Effective Deployment: Engineered to run efficiently on consumer-grade GPUs, such as an RTX 4090D, making it accessible for practical applications.
- Reasoning Structure: Utilizes explicit identifiers for thinking and summary parts (e.g.,
<think> and <answer>) to structure its problem-solving process.
Limitations
- Scenario Specificity: Optimized exclusively for K-12 mathematics; performance in non-mathematical scenarios is not guaranteed.
- Invalid Results: May occasionally produce invalid results due to circular reasoning or parsing issues when explicit thinking/summary tags are used.
- Safety and Ethics: Has not undergone specific optimization or testing for safety and ethical alignment.