netease-youdao/Confucius-o1-14B

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:14.8BQuant:FP8Ctx Length:32kPublished:Jan 20, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Warm

Confucius-o1-14B is a 14.8 billion parameter reasoning model developed by the NetEase Youdao Team, based on the Qwen2.5-14B-Instruct architecture. It employs a two-stage learning strategy to achieve o1-like thinking abilities, specifically optimized for generating a summarized step-by-step problem-solving process from its chain of thought. This model is designed for educational applications, particularly in K12 mathematics, and can be deployed on a single GPU without quantization, supporting a 131072 token context length.

Loading preview...

Confucius-o1-14B: Specialized Reasoning for Education

Confucius-o1-14B is a 14.8 billion parameter model from the NetEase Youdao Team, built upon the Qwen2.5-14B-Instruct base. Its core innovation lies in its o1-like reasoning capabilities and a unique output format that provides both a detailed chain of thought and a concise, step-by-step problem-solving summary.

Key Capabilities & Features

  • Two-Stage Learning: Utilizes a teacher model for initial learning of o1 thinking patterns, followed by self-iterative learning to enhance reasoning.
  • Summarized Problem-Solving: Generates a <thinking> block for detailed reasoning and a <summary> block for a 3-5 step problem-solving process, making complex solutions accessible.
  • Educational Focus: Specifically optimized for K12 mathematics scenarios, with training data sampled from non-graphical math problems.
  • Efficient Deployment: Designed to run on a single GPU without quantization, making it accessible for various applications.
  • Rigorous Data Filtering: Employs automated evaluation to ensure the correctness of both the final answer and the explanation process, preventing false positives.

Use Cases & Limitations

This model is particularly well-suited for educational applications requiring structured mathematical problem-solving explanations. Its primary strength is in K12 mathematics, where it can provide clear, step-by-step solutions. However, users should be aware of its limitations:

  • Scenario Specificity: Performance is verified only in K12 mathematics; quality in other domains is not guaranteed.
  • Language Tendencies: The summary block may lean towards Chinese content, and the thinking block might exhibit mixed language use.
  • Potential for Invalid Results: The model can occasionally fall into circular reasoning, leading to unparseable outputs.
  • No Safety Alignment: The model has not undergone specific safety and ethical alignment testing.

Finetunes
2 models