CondMedQA

A diagnostic benchmark of 100 condition-dependent biomedical questions where the correct answer changes based on patient-specific factors such as comorbidities, allergies, pregnancy status, and contraindications.

Dataset Description

Real-world clinical reasoning is inherently conditional: treatment decisions depend on patient-specific factors such as comorbidities, drug allergies, pregnancy, and organ function. Existing biomedical QA benchmarks largely ignore this conditionality, testing only context-free factual recall.

CondMedQA fills this gap with 100 expert-curated questions that require reasoning over two Wikipedia source articles to arrive at a condition-appropriate answer. Each question presents a clinical scenario where the standard-of-care answer is modified by a specific patient condition.

Key Properties

• Domain: Biomedical / clinical pharmacology / diagnostic imaging
• Question type: Short-answer, condition-dependent
• Multi-hop: Every question requires evidence from 2 source articles
• Total questions: 100
• Answer length: Typically 1–5 words
• Split: Test only (diagnostic benchmark)

Dataset Format

A single CSV file (condmedqa.csv) with the following columns:

Column	Description
id	Question identifier (1–100)
wiki1	URL of the first Wikipedia source article (primary condition/disease)
wiki2	URL of the second Wikipedia source article (drug/test/modifier)
q	The clinical question with patient-specific conditions
a	The condition-appropriate answer

Usage

from datasets import load_dataset

dataset = load_dataset("jashparekh/CondMedQA")

Or directly:

import pandas as pd

df = pd.read_csv("condmedqa.csv")

Intended Use

• Evaluating LLMs on condition-dependent clinical reasoning
• Benchmarking retrieval-augmented generation (RAG) pipelines on context-sensitive medical QA
• Testing whether models can integrate patient-specific constraints with medical knowledge
• Diagnostic evaluation (100 questions — not intended as a training set)

Citation

@article{parekh2026conditiongated,
  title={Condition-Gated Reasoning for Context-Dependent Biomedical Question Answering},
  author={Parekh, Jash Rajesh and Kweon, Wonbin and Chan, Joey and Islamaj, Rezarta and Leaman, Robert and Jiang, Pengcheng and Wei, Chih-Hsuan and Wang, Zhizheng and Lu, Zhiyong and Han, Jiawei},
  journal={arXiv preprint arXiv:2602.17911},
  year={2026}
}

License

MIT