Introduction to MedFailBench
MedFailBench is a synthetic benchmark built by clinicians to evaluate the safety boundaries of medical AI, rather than merely assessing whether a model can provide the correct answer. It poses a new question: which safety boundary has failed?
Key Features
- Error Labeling: The benchmark labels medical AI errors by severity (1-5) and safety gate types, such as missed urgent escalation, unsafe remote dosing, unsafe discharge reassurance, evidence fabrication, unsafe protocol execution, and source support gap.
- Version Release: The current public release (v0.2.1) includes 44 clinician-reviewed synthetic cases with severity annotations.
- Live Leaderboard: It features a live leaderboard preview on HuggingFace.
- Safety Gate Taxonomy: A taxonomy for safety gates and a clinical severity rubric are provided.
- Automated Pipeline: An automated pipeline is available for archiving model-response screening runs.
Data Privacy
It's important to note that MedFailBench does not include any patient data, clinical validation claims, or model rankings.
Licensing Information
MedFailBench is released under Apache-2.0 and CC-BY-4.0 licenses and carries the Zenodo DOI 10.5281/zenodo.21205535.
Blogger's Review: The introduction of MedFailBench offers a fresh perspective on safety assessment in medical AI, emphasizing an in-depth analysis of model error types. This approach not only aids in enhancing the reliability of medical AI but also provides valuable benchmarks and data support for future research.