Mulberry-260k
收藏资源简介:
Mulberry-260k是由南洋理工大学、清华大学、百度公司和中山大学联合创建的多模态学习推理和反思数据集,包含260,000条数据。每条数据都包含一个问题的推理节点树,这些节点丰富、明确且定义良好。数据集通过集体蒙特卡洛树搜索(CoMCTS)方法构建,旨在训练具有逐步推理和反思能力的多模态大语言模型(MLLM)。该数据集的应用领域主要集中在提升MLLM在复杂任务中的推理能力,尤其是在需要深入推理的领域,如数学和视觉语言理解。
Mulberry-260k is a multimodal learning, reasoning and reflection dataset jointly developed by Nanyang Technological University, Tsinghua University, Baidu Inc., and Sun Yat-sen University, which contains 260,000 data entries. Each entry includes a reasoning node tree for a specific question, where the nodes are abundant, explicit and well-defined. The dataset is constructed using the Collective Monte Carlo Tree Search (CoMCTS) method, and is designed to train multimodal large language models (MLLMs) with step-by-step reasoning and reflection capabilities. Its application scenarios primarily focus on enhancing the reasoning performance of MLLMs in complex tasks, especially in domains that demand in-depth reasoning such as mathematics and vision-language understanding.
Mulberry 数据集概述
数据集名称
Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search
数据集发布信息
- 发布日期: 2024年12月24日
- 发布平台: arXiv
- 论文链接: https://arxiv.org/abs/2412.18319
数据集作者
- 主要作者:
- Huanjin Yao (清华大学, 百度)
- Jiaxing Huang (南洋理工大学, 通讯作者)
- Wenhao Wu (百度)
- Jingyi Zhang (南洋理工大学)
- Yibo Wang (清华大学)
- Shunyu Liu (南洋理工大学)
- Yingjie Wang (南洋理工大学)
- Yuxin Song (百度)
- Haocheng Feng (百度)
- Li Shen (中山大学)
- Dacheng Tao (南洋理工大学)
数据集贡献机构
- 南洋理工大学
- 清华大学
- 百度
- 中山大学
数据集简介
Mulberry 数据集旨在通过集体蒙特卡洛树搜索(Collective Monte Carlo Tree Search)增强多模态大语言模型(MLLM)的推理和反思能力,使其具备类似o1的推理能力。

- 1Mulberry: Empowering MLLM with o1-like Reasoning and Reflection via Collective Monte Carlo Tree Search南洋理工大学, 清华大学, 百度公司, 中山大学 · 2024年



