Visual Question Answering (VQA) over charts and tables requires joint reasoning over language, visual structure, and numerical information, yet no standard benchmark currently exists for Vietnamese. We introduce ChartTableViVQA, a Vietnamese VQA dataset with 13,003 question–answer–instruction triplets over 868 chart and table images. Questions cover multiple reasoning types, ranging from simple value lookup to multi-step aggregation, comparison, and trend analysis, and are accompanied by concise instructions that explicitly guide the reasoning process. The dataset is built via a semi-automatic human-in-the-loop pipeline, combining large language model generation with manual verification by Vietnamese annotators. We evaluate several vision-language models in zero-shot and instruction-based settings. Results show strong performance on simple questions but clear weaknesses in complex reasoning, particularly on charts, while instruction-based prompting consistently improves accuracy.