ChartNet-Bench is a benchmark dataset containing 3,807 chart images for evaluating faithful multimodal chart understanding. It includes 2,000 synthetic charts and 1,807 real-world charts, all human-verified. The benchmark supports tasks like chart-to-CSV extraction, summarization, and hallucination detection.
Use Cases
- Evaluate chart-to-CSV extraction performance based on chart images
- Benchmark chart summarization models based on multimodal chart content
- Detect and correct hallucinations in chart analysis outputs based on evidence-grounded evaluation
- Train models for faithful chart understanding based on human-verified instances
Strengths
- Contains 3,807 chart images, providing a substantial testbed
- Includes 2,000 synthetic charts and 1,807 real-world charts, offering diverse sources
- All instances are human-verified, ensuring a quality benchmark
Limitations
- Column-level documentation is absent; field semantics must be inferred after download
- Row count is unknown, which may limit suitability assessment
Provenance
- Source
- Harvard Dataverse
- Freshness
- Last updated 2026-05-06 04:18:59; freshness should be verified