WeThink
收藏资源简介:
WeThink数据集是一个包含超过12万个多模态问答对的数据集,每个问答对都带有明确的推理路径。数据集由18个不同的公开图像数据集精心挑选而来,涵盖了各种问题和领域,包括推理、OCR、识别、数学、知识和空间感知等,从而增强了模型的多模态推理能力。WeThink数据集旨在解决多模态推理中缺乏多样化和推理中心数据的问题,并为模型提供了一组全面且具有挑战性的数据,以帮助模型在视觉语言推理任务中取得更好的表现。
The WeThink dataset is a collection of over 120,000 multimodal question-answer pairs, each paired with explicit reasoning paths. It is carefully curated from 18 distinct public image datasets, covering a wide range of question types and domains including reasoning, OCR, recognition, mathematics, general knowledge, and spatial perception, to enhance the multimodal reasoning capabilities of models. The WeThink dataset aims to address the scarcity of diverse, reasoning-focused data in multimodal reasoning, and provides a comprehensive and challenging dataset to help models achieve better performance on visual-language reasoning tasks.
WeThink数据集概述
数据集基本信息
- 名称: WeThink Multimodal Reasoning 120K
- 类型: 多模态推理数据集(视觉-语言)
- 存储位置: Hugging Face平台
- 数据集地址: https://huggingface.co/datasets/yangjie-cv/WeThink_Multimodal_Reasoning_120K
数据集内容
- 数据格式: JSON
- 字段说明:
problem: 问题描述answer: 答案category: 问题类型abilities: 问题所需能力refined_cot: 思维过程image_path: 图像路径reward: 奖励方法
配套资源
- 配套图像:
- 下载地址: https://huggingface.co/datasets/Xkev/LLaVA-CoT-100k
- 文件格式: 分卷压缩包(image.zip.part-aa至image.zip.part-ap)
- 合并命令:
cat image.zip.part-* > image.zip && unzip image.zip
相关模型
- 关联模型: WeThink-Qwen2.5VL-7B
- 模型地址: https://huggingface.co/yangjie-cv/WeThink-Qwen2.5VL-7B
- 性能表现:
- OpenCompass多模态推理排行榜第1名
- OpenCompass多模态学术排行榜第5名(截至2025年5月30日)
训练支持
- 监督微调推荐工具: LLaMA-Factory
- 强化学习框架: EasyR1(支持基于规则的强化学习)
引用信息
bibtex @misc{yang2025wethink, title={WeThink: Toward General-purpose Vision-Language Reasoning via Reinforcement Learning}, author={Jie Yang and Feipeng Ma and Zitian Wang and Dacheng Yin and Kang Rong and Fengyun Rao and Ruimao Zhang}, year={2025}, eprint={2506.07905}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2506.07905}, }

- 1WeThink: Toward General-purpose Vision-Language Reasoning via Reinforcement Learning腾讯微信视觉实验室 · 2025年



