EpiInfDisQA

EpiInfDisQA is an infectious-disease-focused question-answering dataset assembled by filtering three existing open medical QA benchmarks down to their infectious-disease-relevant subsets and normalizing them into a single schema.

Dataset Summary

71,725 question/answer records covering infectious disease topics — exam-style multiple-choice questions (USMLE, Indian medical entrance exams) and research-question QA over PubMed abstracts — drawn from:

Source	License	Kept / Scanned
MedQA-USMLE	CC-BY-4.0	2,428 / 11,451
MedMCQA	Apache-2.0	37,256 / 187,005
PubMedQA (pqa_labeled + pqa_artificial)	MIT	32,041 / 212,269

Records were kept if their question, answer, or topic text matched an infectious-disease keyword or acronym list (pathogens, disease names, antimicrobials, vaccination/outbreak terms, etc. — see Limitations below for what this means for precision).

Dataset Structure

Data Fields

• id (string): unique record id, prefixed by source (e.g. medqa_train_18).
• source (string): one of medqa, medmcqa, pubmedqa.
• topic (string): subject/topic label — MedMCQA's subject_name (e.g. "Microbiology", "Surgery"), MedQA's exam step (e.g. "USMLE (step1)"), or PubMedQA's top MeSH terms.
• date (null): always null — none of the three source datasets carry a per-question or per-article publication date.
• question (string): the question text. For multiple-choice sources (MedQA, MedMCQA) the options are folded in as A) ... B) ... C) ... D) .... For PubMedQA, the source abstract is appended as Context: ... so the question is self-contained.
• answer (string): the correct answer. For MCQ sources, the correct option's letter + text (plus an Explanation: block, when MedMCQA provides one). For PubMedQA, the paper's free-text conclusion (long_answer).
• source_url (string): the source dataset's Hugging Face page (no per-question URL exists upstream).

Data Splits

Single train split, 71,725 records. No held-out split is provided — consumers who need train/val/test should split downstream (the underlying MedQA/MedMCQA test-labeled splits are already folded into this single file).

Limitations

• The infectious-disease filter is keyword-based and recall-oriented, not a ground-truth label from the source datasets. It will include false positives — e.g. a paper about pediatric ward efficiency that happens to mention "pneumonia" and "gastroenteritis" as common admission diagnoses, without infectious disease being its actual subject. Spot-check before using for training, and tighten the keyword list if you need higher precision.
• date is always null. These are exam question banks and biomedical abstracts, not dated news/reference content — there is no reliable per-record date to attach (PubMedQA's pubid could in principle be resolved to a real publication date via the PubMed API, but this hasn't been done here).
• MedMCQA's official test split is excluded — its cop (correct option) field is unlabeled (-1) on that split, so only train + validation are usable for QA purposes.
• IDQuAD, a purpose-built infectious-disease QA dataset from a 2025 paper (Yonsei/Yarmouk universities), was considered but has no public Hugging Face release — only the paper and its supplementary files — so it is not included here.

Licensing Information

This dataset is a derivative work combining three sources under different licenses:

• MedQA-USMLE: CC-BY-4.0
• MedMCQA: Apache-2.0
• PubMedQA: MIT

The combined dataset is released under CC-BY-4.0 — the most restrictive of the three (it requires attribution), applied uniformly since license information is not tracked per-row. Attribution to all three original sources is required per their license terms; see the citations below.

Citation Information

@article{jin2021disease,
  title={What disease does this patient have? a large-scale open domain question answering dataset from medical exams},
  author={Jin, Di and Pan, Eileen and Oufattole, Nassim and Weng, Wei-Hung and Fang, Hanyi and Szolovits, Peter},
  journal={Applied Sciences},
  volume={11},
  number={14},
  pages={6421},
  year={2021},
  publisher={MDPI}
}

@inproceedings{pal2022medmcqa,
  title={MedMCQA: A Large-scale Multi-Subject Multi-Choice Dataset for Medical domain Question Answering},
  author={Pal, Ankit and Umapathi, Logesh Kumar and Sankarasubbu, Malaikannan},
  booktitle={Conference on Health, Inference, and Learning},
  pages={248--260},
  year={2022},
  organization={PMLR}
}

@inproceedings{jin2019pubmedqa,
  title={PubMedQA: A Dataset for Biomedical Research Question Answering},
  author={Jin, Qiao and Dhingra, Bhuwan and Liu, Zhengping and Cohen, William and Lu, Xinghua},
  booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)},
  pages={2567--2577},
  year={2019}
}