MIT and MIT-IBM Computing Research Lab researchers developed ChartNet, a synthetic dataset of over one million chart images designed to train vision-language models (VLMs) for chart interpretation tasks. The dataset pairs chart images with generation code, textual descriptions, numerical tables, and Q&A pairs to enable robust multimodal reasoning. A two-step pipeline translates existing charts into code, then iteratively augments them across chart types, colors, and data values. Models trained on ChartNet showed significant improvements in chart reconstruction, data extraction, summarization, and question answering — with smaller open-source models outperforming much larger commercial counterparts. The dataset is open-source and aims to democratize chart-understanding AI for businesses with limited budgets.
