VivaBench: 1,152 Clinical Vignettes for Simulating Oral Exams

1,152 physician-curated clinical vignettes simulate multi-turn viva voce (oral) examinations. The dataset, VivaBench, is designed to evaluate clinical reasoning in large language models by requiring agents to iteratively gather history and physical findings and order investigations. It was created by author chychiu and last updated on October 9, 2025.

Use Cases

• Benchmarking LLM diagnostic accuracy based on multi-turn clinical reasoning tasks
• Training AI agents for interactive patient history gathering based on simulated exam scenarios
• Evaluating the step-by-step clinical reasoning process of AI models based on iterative questioning

Strengths

• Contains 1,152 clinical vignettes curated by physicians
• Simulates a multi-turn, interactive viva voce examination format

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment

Provenance

Source

huggingface

Collection Method

Physician-curated clinical vignettes

Freshness

Last updated 2025-10-09 11:31:48; freshness should be verified