ViChartVQA-Dataset: A Cost-Efficient AI-Generated Vietnamese Chart Visual Question Answering Dataset

ViChartVQA-Dataset is a Vietnamese Chart Visual Question Answering dataset comprising 4,180 chart-question-answer triplets across four chart types (bar, line, area, pie). The dataset is generated through a cost-efficient AI-assisted pipeline that replaces expensive vision-language API calls with a text-only LLM workflow: chart data is rendered locally via matplotlib and exported as compact Markdown tables, which are then sent to GPT-4o-mini for question and answer generation. This approach reduces API costs by 10–20× compared to conventional vision-based pipelines. Human evaluation by three Vietnamese native speakers confirms dataset quality with Fleiss' Kappa κ = 0.68–0.74. The dataset is split 80/10/10 (train/validation/test) by chart ID to prevent data leakage.