Counterfactual Pendulum Multilingual: Visual-Textual Conflict Scenarios for VLM Evaluation

A dataset designed to test Vision-Language Model behavior when presented with conflicting visual and textual information. It adapts the Counterfactual Pendulum scenario across two visual conflict dimensions: Angular (Angle) and Light. The dataset was created by apart-global-south-hack and was last updated on June 20, 2026.

Use Cases

• Evaluate VLM reliance on visual evidence versus textual bias based on contradictory pendulum angle information.
• Benchmark model honesty or abstention rates when faced with unresolvable conflict in lighting conditions.
• Study multimodal reasoning failures in models using controlled visual-textual conflict scenarios.
• Develop methods to mitigate textual bias in VLMs using the described conflict dimensions.

Strengths

• Focuses on a specific, controlled evaluation scenario for Vision-Language Models.
• Explicitly tests two distinct conflict dimensions: Angular and Light.
• Dataset was last updated on 2026-06-20, suggesting recent maintenance.

Limitations

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

Provenance

Source

huggingface

Freshness

Last updated 2026-06-20 11:57:25; freshness should be verified.