MIT, IBM Researchers Build 1 Million-Chart Dataset, Lifting Smaller AI Models Past Commercial Rivals
Updated
Updated · MIT News · Jun 3
MIT, IBM Researchers Build 1 Million-Chart Dataset, Lifting Smaller AI Models Past Commercial Rivals
1 articles · Updated · MIT News · Jun 3
More than 1 million synthetic chart images power ChartNet, a new MIT and MIT-IBM dataset built to train vision-language models to extract, summarize and answer questions about charts.
A two-step pipeline converts existing charts into code, then generates hundreds of variations while attaching text descriptions, numerical tables and Q&A pairs to align visual, linguistic and numeric reasoning.
Tests on IBM’s Granite Vision models and other open-source systems showed accuracy gains across chart reconstruction, data extraction, summarization and question answering, with smaller models beating much larger commercial rivals.
Human-annotated samples are included for fine-tuning specialized applications, and the open-source release could help smaller firms use AI for business trend analysis and scientific figure interpretation at lower cost.
The team plans to expand ChartNet with more complex data and present the work at the IEEE Computer Vision and Pattern Recognition Conference.
Can AI trained on perfect synthetic data master the messy, imperfect charts of the real world?
If smarter data lets small AI beat tech giants, is the era of massive AI models already ending?
When powerful AI analysis is available to everyone, who truly gains the competitive edge in business?
ChartNet: Scaling Chart Interpretation with 1M+ Synthetic and Real-World Annotated Visualizations
Overview
ChartNet, developed by MIT and the MIT-IBM Computing Research Lab, addresses the long-standing challenge of limited and low-quality training data for chart interpretation by introducing a massive, richly annotated dataset of over one million synthetic chart images. This breakthrough enables smaller, open-source AI models to perform advanced chart interpretation tasks with high accuracy, a feat previously restricted by data scarcity. By providing detailed annotations, generating code, and question-answer pairs, ChartNet empowers robust training of Vision-Language Models, making sophisticated chart understanding more accessible and efficient for a wider range of researchers and applications.