库名称：transformers
许可证：其他
基础模型：deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
标签：

• llama-factory
• full
• generated_from_trainer
  模型索引：
• 名称：ReasonFlux-F1-32B
  结果：[]

ReasonFlux：基于思维模板扩展的分层大语言模型推理

革命性的模板增强推理范式使32B模型在推理任务中超越o1-mini和DeepSeek-R1蒸馏模型。

ReasonFlux-F1-32B

ReasonFlux-F1-32B是我们通过利用ReasonFlux-Zero中的模板增强推理轨迹微调出的SOTA级推理大语言模型。

• GitHub仓库：Gen-Verse/ReasonFlux
• 论文：ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates
• 数据集：Gen-Verse/ReasonFlux-F1-SFT

评估

我们在AIME2024、AIME2025、MATH500和GPQA-Diamond等具有挑战性的推理任务上展示了ReasonFlux-F1-32B的评估结果。为确保公平比较，我们在ReasonFlux-F1中报告了各模型的评估脚本结果。

使用VLLM快速开始

from vllm import LLM, SamplingParams  
from transformers import AutoTokenizer  

model_id = 'Gen-Verse/ReasonFlux-F1'  

model = LLM(  
    model_id,  
    tensor_parallel_size=8,  
)  
tokenizer = AutoTokenizer.from_pretrained(model_id)  

sampling_params = SamplingParams(  
    max_tokens=32768,  
)  
# 2022 AIME I Problems/Problem 15  
question = """设\(x, y\)和\(z\)为正实数，满足以下方程组：  
\[  
\begin{array}{c}  
\sqrt{2 x-x y}+\sqrt{2 y-x y}=1 \\  
\sqrt{2 y-y z}+\sqrt{2 z-y z}=\sqrt{2} \\  
\sqrt{2 z-z x}+\sqrt{2 x-z x}=\sqrt{3} .  
\end{array}  
\]  
则\(\left[(1-x)(1-y)(1-z)\right]^{2}\)可表示为\(\frac{m}{n}\)，其中\(m\)和\(n\)为互质的正整数。求\(m+n\)。"""  
ds_prompt="\n" + question + "\n"  
output = model.generate(ds_prompt, sampling_params=sampling_params)  
print(output[0].outputs[0].text)  

引用

@article{yang2025reasonflux,  
  title={ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates},  
  author={Yang, Ling and Yu, Zhaochen and Cui, Bin and Wang, Mengdi},  
  journal={arXiv preprint arXiv:2502.06772},  
  year={2025}  
}