MedQA-MM Identifier Release

Paper repository · Hugging Face dataset

MedQA-MM is a 1,000-item shortcut-mitigated medical multimodal multiple-choice benchmark constructed from MedThinkVQA, MedXpertQA-MM, and the Health and Medicine portion of MMMU. This public release is intentionally identifier-only.

It does not contain source questions, answer choices, gold answers, images, clinical text, or repaired payloads. It provides stable source locators, pinned source revisions, and a reconstruction script so that users who have independently obtained the source datasets under their respective terms can reproduce a local evaluation copy.

Released files

• data/medqa_mm_ids.jsonl: one viewer-friendly locator record per MEDQA-MM item.
• medqa_mm_ids.json: the original structured 1,000-record locator release and validation metadata.
• source_revisions.json: official source repositories, pinned revisions, subsets, and terms links.
• reconstruct_medqa_mm.py: validates all locators and optionally materializes a local source-linked evaluation copy.
• requirements.txt: minimal reconstruction dependencies.

Composition

Source dataset	Eligible candidate items	MEDQA-MM items
MedThinkVQA	720	207
MedXpertQA-MM	2,000	450
MMMU Health/Medicine	1,712	343
Total	4,432	1,000

For MMMU Health/Medicine, the five source test subsets contain 1,752 rows; the 1,712 candidate count excludes 40 open-response rows and matches the multiple-choice eligibility rule used by the construction pipeline.

Source repositories and terms

MEDQA-MM source	Official repository	Pinned revision	Terms relevant to reconstruction
MedThinkVQA	bio-nlp-umass/MedThinkVQA	5555b80e167c4796efab92fe709fb05914077dc0	CC BY-NC-SA 4.0; research and education only; source card restrictions apply.
MedXpertQA-MM	TsinghuaC3I/MedXpertQA	7e7c465a68eb2b866926bfa59c8c9d17a8daba65	The paper states evaluation-only use and asks users not to share examples online. Do not redistribute the reconstructed payload.
MMMU Health/Medicine	MMMU/MMMU	98e6ac0cb9b7b2cd2c991b85a50762edc4aedc68	Apache-2.0 dataset card; source-level copyright and removal disclaimer also applies.

The upstream sources retain all rights in their questions, answers, images, and associated content. Publishing this locator list does not grant permission to redistribute any source payload.

Locator schema

Each JSONL row contains:

• medqa_mm_id: stable MEDQA-MM identifier.
• dataset: one of MedThinkVQA, MedXpertQA, or MMMU-HM.
• source_repo and source_revision: official pinned source.
• source_config, split, example_id, and row_index: primary reconstruction locator.
• case_number or mmmu_subset where needed for fallback matching.
• source_license_or_terms and source_terms_url: source-specific release notice.

The original JSON additionally records the validation performed against the complete construction manifests.

Validate the identifiers

From this directory:

python -m pip install -r requirements.txt
python reconstruct_medqa_mm.py \
  --output-dir medqa_mm_validation

The default mode downloads only the source metadata needed to resolve the locators. It writes:

• resolved_ids.jsonl
• validation_summary.json

Successful validation requires all 1,000 locators to match exactly and reproduces the 207/450/343 source counts.

Create a local evaluation copy

Read all three source terms first. Then run:

python reconstruct_medqa_mm.py \
  --output-dir medqa_mm_local \
  --materialize \
  --accept-source-terms

This additionally creates:

• medqa_mm.jsonl: standardized local MCQ records.
• images/MEDQA-MM-XXXX/: locally materialized source images.

The output is for local evaluation under the upstream terms. Do not upload or redistribute it. It is a source-linked reconstruction and does not include unpublished repair artifacts that are not part of this identifier release.

License and responsible use

This repository uses license: other because the release combines author-created locator metadata with references to independently governed source datasets. No license to the underlying source payloads is granted here.

MedQA-MM is a research benchmark, not a clinical device. It must not be used for diagnosis, treatment, triage, or other clinical decision-making. Users are responsible for reviewing the current upstream licenses, terms, privacy notices, and source revisions before reconstruction.

Citation

@inproceedings{wang2026medqamm,
  title     = {MedQA-MM: Shortcuts Behind Medical Visual Reasoning},
  author    = {Wang, Benlu and Zhang, Yifan and Yu, Jiaqing and Ong, Chin Siang and Huang, Juncheng and Li, Zhuohao and Zhang, Zhenyu and Cohan, Arman and Yu, Hong and Yao, Zonghai},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}

Please also cite the three original source datasets when using a reconstructed local copy.