VietTravelVQA

Objective: This dataset is designed to benchmark and improve Visual Question Answering (VQA) systems in the context of Vietnamese tourism and cultural heritage. It addresses the lack of high-quality, regionally specific multimodal data for Southeast Asia. Data Content: The dataset comprises thousands of images sourced from Wikimedia Commons, paired with human-verified question-answer sets in Vietnamese. The questions cover five levels of complexity, ranging from basic object identification to deep cultural reasoning. Methodology: 1. Sourcing: Legally compliant images were filtered from Wikimedia Commons. 2. Annotation: Expert annotators generated QA pairs, focusing on architectural details, historical significance, and spatial reasoning. 3. Validation: Data was cleaned using automated scripts to ensure 100% synchronization between metadata (JSON) and image files, with factual auditing via Large Multimodal Models (LMMs). Usage: The data is split into train and test sets (.json). It is intended for training, fine-tuning, and evaluating Vision-Language Models (VLMs) on localized cultural contexts.任务目标：本数据集旨在针对越南旅游与文化遗产场景下的视觉问答（Visual Question Answering, VQA）系统开展基准测试与性能优化，以解决东南亚地区缺乏高质量、区域专属多模态数据的行业痛点。 数据内容：本数据集包含数千张源自维基共享资源（Wikimedia Commons）的图像，并搭配经人工校验的越南语问答对。问题覆盖五个复杂度层级，从基础的物体识别任务逐步延伸至深层次的文化推理任务。 研究方法： 1. 数据采集：从维基共享资源（Wikimedia Commons）筛选合规合法的图像资源。 2. 标注流程：由专业标注人员生成问答对，重点聚焦建筑细节、历史意义与空间推理维度。 3. 数据验证：通过自动化脚本完成数据清洗工作，确保元数据（JSON格式）与图像文件完全同步，并借助大型多模态模型（Large Multimodal Models, LMMs）开展事实性审核。 使用场景：本数据集已划分为训练集与测试集（均为JSON格式），可用于面向本地化文化场景的视觉语言模型（Vision-Language Models, VLMs）的训练、微调与评估工作。