VisCoR-55K: Contrastive Visual Question-Answering Pairs

VisCoR-55K contains contrastive visual question-answering pairs spanning five categories: General, Reasoning, Math, Graph/Chart, and OCR. The dataset was created by author 5551z and was last updated on HuggingFace in April 2026. It includes original VQA samples, matched contrastive counterparts, and generated rationales.

Use Cases

• Training models for faithful visual reasoning based on the described contrastive pairs.
• Benchmarking visual question-answering systems across the five described categories.
• Developing contrastive learning techniques for multimodal AI based on the structured pairs.
• Analyzing model robustness using the constructed counterfactual examples.

Strengths

• Dataset spans five distinct categories of visual reasoning tasks.
• Contains matched contrastive pairs specifically constructed to encourage faithful reasoning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

5551z on HuggingFace

Freshness

Last updated 2026-04-30 01:37:13.