README.md

metadata

license: cc-by-sa-4.0
task_categories:
  - multiple-choice
  - question-answering
language:
  - de
pretty_name: Clinical Practice Guideline Benchmark Questions in German
size_categories:
  - n<1K

Dataset Card for cpgQA-DE

cpgQA-DE is a German benchmark for evaluating whether medical question-answering systems can answer questions grounded in German Clinical Practice Guidelines (CPGs).

Evaluation only. Do not use for training.

This is a research benchmark, not a clinical decision-support tool.

Dataset Details

This repository provides four evaluation files.

cpgQA_DE_main.json

Contains 200 validated guideline-based questions with answer keys, source references, contextual excerpts, and difficulty/relevance annotations.

Fields:

• question_id: Unique question identifier
• guideline_title: Title or identifier of the clinical practice guideline
• question_stem: Main question text
• correct_answer: Correct answer option
• distractor_1 to distractor_5: Incorrect answer options; distractor_5 is "I don't know"
• context_excerpt: Supporting excerpt from the guideline text
• guideline_page: Page number in the guideline
• comment: Additional note or annotation
• context_type: Context source type (figure, table, full_text, recommendation_box)
• question_type: Question type, such as diagnostic_whats_next, treatment_whats_next, knowledge, or error_detection
• guideline_url: URL of the source guideline
• modified_after_validation: Whether the question was modified after expert validation
• difficulty_resident: Perceived difficulty for a resident physician
• difficulty_specialist: Perceived difficulty for a board-certified specialist
• clinical_relevance: Perceived clinical importance
• requires_recall_or_lookup: Whether a physician should be able to answer without external resources
• frequency_in_practice: Estimated frequency of the topic in clinical practice

cpgQA_DE_4opt.json

Simplified 4-option multiple-choice format for LLM or RAG evaluation.

Fields:

• ID: Question identifier
• question: Question stem
• A: First answer option
• B: Second answer option
• C: Third answer option
• D: Forth answer option
• answer: Correct answer label (for example “A”)

cpgQA_DE_5opt.json

Extends the 4-option format with option E: "The question is flawed; no correct answer possible." It is used to test error detection and uncertainty calibration in language models.

cpgQA_DE_6opt.json

Extends the 5-option format with option F: "I don't know."

Dataset Authors

Curated by Johannes Schwietering and Gregor Lichtner.

License: CC BY-SA 4.0.

Citation

If you use cpgQA-DE in your research, please cite the accompanying paper:

Schwietering J, Lichtner G. Benchmarking large language models for question answering on German clinical practice guidelines. BMJ Health & Care Informatics. 2026;33:e102207. https://doi.org/10.1136/bmjhci-2026-102207

BibTeX:

@article{schwietering_cpgqade_2026,
  title   = {Benchmarking Large Language Models for Question Answering on German Clinical Practice Guidelines},
  author  = {Schwietering, Johannes and Lichtner, Gregor},
  journal = {BMJ Health \& Care Informatics},
  year    = {2026},
  volume  = {33},
  pages   = {e102207},
  doi     = {10.1136/bmjhci-2026-102207},
  url     = {https://doi.org/10.1136/bmjhci-2026-102207}
}