BannedDrug-Bench

Dataset Summary

BannedDrug-Bench is an adversarial evaluation framework designed to test clinical safety regression in Large Language Models (LLMs). The dataset contains 103 clinical Multiple Choice Questions (MCQs) created using adverse drug data available from DrugBank (https://go.drugbank.com) where the traditionally CORRECT options are pharmaceuticals that have been withdrawn, banned, or severely restricted by global regulatory bodies (FDA, NIH, EMA).

This benchmark tests whether a model relies on its static, outdated training data (parametric memory) or if it can successfully trigger a safe refusal when faced with a clinical safety trap.

Observational Data & Interaction Logs

To support the findings in "Trust but Verify," the following persistent interaction logs document the clinical hallucinations and safety regressions identified during Experiment III. These transcripts provide a raw look at how various frontier models handled the adversarial prompts.

Model Reference	Interaction Log Link
GPT-5.1	View Interaction Log
GPT-5.3	View Interaction Log
Gemini 3 Flash	View Interaction Log
Gemini 3.1 Pro	View Interaction Log

Dataset Structure

• Sample Count: 103 Adversarial MCQs
• Language: English
• Format: JSON
• Key Features:
  • question: Clinical case vignette or pharmacological query.
  • options: 4 choices (A, B, C, D) including the banned target drug.
  • correct_option: The historically accurate but now prohibited selection.
  • regulatory_status: Specific reason for withdrawal (e.g., Hepatotoxicity, Cardiac Arrest).
  • verification_source: Ground-truth evidence from DrugBank (https://go.drugbank.com) and official FDA/NIH archives.

Ethical & Safety Statement

BannedDrug-Bench is an adversarial dataset for AI Safety Research. It contains information about withdrawn and potentially lethal medical practices. It is strictly intended for benchmarking clinical decision support systems and must never be used as a source for actual medical advice or to train models to suggest unsafe treatments.

Citation

If you utilize this dataset in your research, please cite:

@mastersthesis{osama2026trust,
  title={Trust but Verify: Mitigating Medical Hallucinations via Post-Hoc Adversarial Auditing},
  author={Muhammad Osama},
  school={SINES NUST},
  year={2026},
  url={[https://huggingface.co/datasets/muhammadocama/BannedDrug-Bench](https://huggingface.co/datasets/muhammadocama/BannedDrug-Bench)}
}