Human Hallucination Verification Dataset for Multimodal Models

HHVD is a Human Hallucination Verification Dataset for multimodal hallucination verifiability. It contains 4,470 time-constrained human responses to image-text pairs, designed to evaluate obvious and elusive hallucinations. The dataset was created by BeEnough and last updated in April 2026.

Use Cases

• Benchmarking model hallucination rates against human verification times from the 4,470 human responses.
• Training classifiers to detect 'obvious' versus 'elusive' hallucinations based on human-labeled verifiability.
• Analyzing the correlation between image-text pair content and the difficulty of human hallucination verification.
• Developing metrics for multimodal hallucination severity grounded in human response time and accuracy data.

Strengths

• Contains 4,470 human-generated verification responses.
• Specifically designed to evaluate both obvious and elusive hallucination types.

Limitations

• Dataset size of 4,470 responses may be limited for training large models.
• Specific demographic or geographic coverage of human evaluators is unknown, which may introduce bias.
• Column details and data structure are not provided in the input.

Provenance

Source

BeEnough (author/organization on Hugging Face).

Collection Method

Constructed from time-constrained human responses to image-text pairs.

Freshness

Last updated on 2026-04-07.