Vero Visual Reasoning Dataset for Multimodal AI Training

Vero-600k is a collection of data for training and evaluating general visual reasoning models, created by researchers at Princeton University's zlab. The dataset supports broad multimodal reasoning tasks across charts, STEM problems, spatial reasoning, and knowledge grounding. It was released in early 2026.

Use Cases

• Training models for visual instruction following using image-text instruction pairs.
• Evaluating spatial reasoning performance on specific visual problem sets.
• Fine-tuning models for STEM question answering from diagrams and charts.
• Developing counting and grounding capabilities using visual examples.
• Benchmarking knowledge-based reasoning across multimodal inputs.

Strengths

• Released by a Princeton University research lab (zlab-princeton).
• Designed for broad multimodal reasoning across 7+ distinct task categories.
• Includes associated training code and evaluation benchmarks for reproducibility.

Limitations

• Unknown total size, row count, and specific data composition.
• Potential bias towards academic or synthetic visual reasoning problems.
• Lacks detailed documentation on data collection and annotation quality.

Provenance

Source

zlab-princeton (Princeton University)

Collection Method

Method of data gathering is not specified in the provided description.

Freshness

Last updated on 2026-04-07.