VisText: Semantic Chart Captioning Benchmark from MIT

Developed by the MIT Visualization Group (mitvis) and updated in 2025, VisText is a benchmark dataset for chart captioning. It provides paired chart images and captions to evaluate how models interpret visual data representations.

Use Cases

• Training T5 models for automated chart-to-text generation
• Evaluating semantic depth in image captioning systems
• Benchmarking visual reasoning for chart images

Strengths

• Developed by the MIT Visualization Group (mitvis)
• GPL-3.0 licensed for open research
• Focuses on semantic data interpretation rather than simple visual descriptions

Limitations

• Unknown record count
• Lack of detailed schema for metadata fields

Provenance

Source

MIT Visualization Group (mitvis)

Freshness

Last updated in August 2025.