DoraVQA
收藏资源简介:
DoraVQA是由东北大学团队构建的教育视频问答数据集,基于《爱探险的朵拉》8季96集的视频内容,包含5,344条严格时间对齐的问答对。该数据集通过自动解析视频字幕和视觉停顿片段,保留了节目中特有的'上下文-提问-停顿-回答'教学结构,涵盖空间定位、导航决策等60.6%的空间推理任务。数据创建过程采用LLM代理自动对齐时间戳并提取教学逻辑,38小时的视频内容包含78.8%即时推理和23.2%序列推理问题。该数据集旨在提升视觉语言模型在空间推理、多模态理解等领域的表现,通过教育视频的结构化监督信号解决现有模型在基础推理任务上的系统性缺陷。
DoraVQA is an educational video question answering dataset constructed by the team from Northeastern University. It is based on 96 episodes across 8 seasons of the TV series *Dora the Explorer*, and contains 5,344 strictly time-aligned question-answer pairs. This dataset preserves the unique "context-question-pause-answer" instructional structure inherent in the original program, which is obtained by automatically parsing video subtitles and visual pause segments. It covers 60.6% of spatial reasoning tasks such as spatial localization and navigation decision-making. During the data creation process, LLM Agents were used to automatically align timestamps and extract instructional logic. The 38 hours of video content includes 78.8% immediate reasoning questions and 23.2% sequential reasoning questions. This dataset aims to improve the performance of vision-language models in fields such as spatial reasoning and multimodal understanding, and addresses the systematic shortcomings of existing models in basic reasoning tasks through structured supervisory signals from educational videos.
DORA-Learning-Spatial-Reasoning 数据集概述
数据集基本信息
- 数据集名称:DORA-Learning-Spatial-Reasoning
- 核心用途:用于教育推理任务,特别是从教育视频中学习空间推理。
- 数据来源:基于动画片《爱探险的朵拉》(Dora the Explorer)的问答视频。
- 关联研究:项目对应研究论文为《Structured Over Scale: Learning Spatial Reasoning from Educational Video》(arXiv:2601.23251)。
数据集内容与构建
- 数据类型:包含视频、字幕(SRT文件)及生成的问答对。
- 构建方式:通过处理视频帧和转录文本上下文来生成数据集,旨在训练模型结合视觉和文本信息提供准确答案。
- 处理脚本:主要使用
scripts/generate_grpo_dataset.py从视频和SRT文件生成数据集。
数据集使用目的
- 模型训练:用于微调QWEN-VL等视觉语言模型,采用GRPO(Group Relative Policy Optimization)方法。
- 任务目标:训练模型回答来自《朵拉》视频的问题,完成教育推理任务。
获取与准备
- 依赖安装:需通过
pip install -r requirements.txt安装项目依赖。 - 生成步骤:
- 运行
scripts/validate_all_seasons.py验证并过滤标签。 - 运行
scripts/generate_grpo_dataset.py脚本,指定标签、视频、SRT文件的目录及输出路径,可处理第1至8季的内容。
- 运行
相关资源
- 论文地址:https://arxiv.org/abs/2601.23251
- 项目主页:https://ostadabbas.github.io/dora.github.io/
- Hugging Face模型:https://huggingface.co/bishoygaloaa/Qween
- Hugging Face Space:https://huggingface.co/spaces/bishoygaloaa/Qween-Space
许可信息
- 许可证:MIT License




