muhammadocama/medverify-dataset

MedVerify

MedVerify is a medical question-answering and regulatory safety validation dataset developed for evaluating large language model (LLM) responses to medical multiple-choice questions, with a focus on post-hoc verification of potentially unsafe, outdated, or regulatory-sensitive information.

This dataset is associated with the manuscript:

Post-Hoc Adversarial Regulatory Verification of Large Language Model Responses to Medical Questions

Dataset Contents

The dataset contains 1,003 medical multiple-choice questions and independent validation responses from MBBS and PharmD reviewers.

The repository contains three files:

File	Description
mcqs.json	Medical multiple-choice questions and associated metadata
mbbs_validator_responses.xlsx	Validation responses from MBBS reviewers
pharmd_validator_responses.xlsx	Validation responses from PharmD reviewers

Dataset Structure

1. Medical MCQs

mcqs.json contains the 1,003 medical multiple-choice questions used in the evaluation.

The questions include different intents, including:

• Historical medical information identification
• Current medical recommendation and safety questions

The questions involving current safety are designed to evaluate whether an LLM can distinguish medically or regulatorily unsafe options from appropriate current information.

2. MBBS Validation

mbbs_validator_responses.xlsx contains validation responses from MBBS-qualified reviewers.

The validation includes assessments of model responses with respect to medical correctness and safety-related characteristics.

3. PharmD Validation

pharmd_validator_responses.xlsx contains validation responses from PharmD-qualified reviewers.

This validation provides an independent pharmacy-focused assessment, particularly relevant to medication safety and regulatory status.

Intended Use

MedVerify is intended for research on:

• Medical LLM evaluation
• Medical hallucination detection
• Regulatory safety verification
• Post-hoc auditing of LLM-generated medical responses
• Medical question-answering benchmarks
• Adversarial evaluation of medical AI systems
• Comparison of different LLM inference and verification strategies

Regulatory Safety

Some questions concern medications or medical information whose regulatory status may have changed over time.

Regulatory status can differ across jurisdictions and may change after the dataset was created. Therefore, this dataset should not be used as a current regulatory or clinical reference.

For clinical or safety-critical applications, information should be independently verified using current authoritative sources, including relevant regulatory authorities, official drug databases, and clinical guidelines.

Limitations

• The dataset represents a finite collection of medical questions and does not cover the full scope of medical knowledge.
• Regulatory classifications may vary between jurisdictions.
• Regulatory and clinical information can change over time.
• Expert validation represents reviewer assessments and should not be interpreted as an absolute clinical ground truth.
• Results obtained using this dataset should not be interpreted as evidence that an LLM is clinically safe or suitable for clinical deployment.
• The dataset is intended for research and evaluation and not for providing medical advice.

Ethical and Safety Considerations

The dataset is designed for evaluating AI safety and does not provide clinical recommendations for patients.

Researchers using the dataset should avoid presenting model performance as evidence of clinical effectiveness or safety without additional prospective evaluation.

License

This dataset is released under the Apache License 2.0.

Contact

For questions, comments, or requests regarding the dataset:

Muhammad Osama

Email: ocama_aslam@outlook.com

Citation

If you use this dataset in your research, please cite the associated manuscript:

@article{osama2026posthoc,
  title   = {Post-Hoc Adversarial Regulatory Verification of Large Language Model Responses to Medical Questions},
  author  = {Osama, Muhammad},
  year    = {2026}
}