BoHanMint/Synthesizer-8B-math

Introduction

CoT-based-Synthesizer-math-8B is a lightweight yet powerful model designed to enhance the reasoning performance of LLMs by synthesizing high-quality answers from multiple candidate responses. It leverages CoT reasoning to analyze and integrate complementary information from flawed or incomplete candidate answers, even when all candidates are incorrect. The model is trained on a large-scale synthetic dataset derived from the original MATH benchmark (12k training samples). Using Llama3.1-70B-Instruct, we generated 294k high-quality synthesized answers.

For further information and details of training, refer to our paper: "CoT-based Synthesizer: Enhancing LLM Performance through Answer Synthesis" available on arXiv.

You can also visit the GitHub repo of the paper and the training dataset used for Synthesizer-8B-math is openly available under Synthesizer-8B-math-train-data.

Model Information

The model is trained based on Llama3.1-8B-Instruct.

Prompt Template

The prompt we used for generating synthesis answer is introduced below.

Please act as an excellent summarizer and summarize the following AI responses to the questions. Your summary should fully consider the connection between the question and AI responses, resulting in a correct, high-quality answer. In most cases, the same response that appears most often in the response may be the correct answer. If you find that there is no correct answer, please try to generate a correct answer yourself. Do not copy The candidate's answer, give your summarized answer and reasons, and give the correct answer at the end of the sentence in the format: The answer is...

[The Start of Original Question]
{question}
[The End of Original Question]

[The Start of AI Responses]
{responses}
[The End of AI Responses]

Citation

If you find our paper helpful, please cite the original paper:

@misc{zhang2025cotbasedsynthesizerenhancingllm,
      title={CoT-based Synthesizer: Enhancing LLM Performance through Answer Synthesis}, 
      author={Bohan Zhang and Xiaokang Zhang and Jing Zhang and Jifan Yu and Sijia Luo and Jie Tang},
      year={2025},
      eprint={2501.01668},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.01668}, 
}

Contact

If you have any questions, we encourage you to either create Github issues or get in touch with us at zbhmint@ruc.edu.cn.