ChartNet-Bench: A Benchmark for Faithful Multimodal Chart Understanding

ChartNet-Bench is a benchmark dataset for faithful multimodal chart understanding. It contains 3,807 chart images, including 2,000 synthetic charts and 1,807 real-world charts. The benchmark supports chart-to-CSV extraction, chart summarization, hallucination detection/correction, and evidence-grounded evaluation. All instances are human-verified and designed to evaluate whether vision-language models can recover, summarize, and verify chart information faithfully.