txz32102/pediamed-mdle-challenge

PediaMed MDLE: Case-Report Challenge Set

This research dataset is a frozen challenge subset associated with Medical Doctor's Last Exam, the PediaMed clinical-reasoning benchmark.

It contains 489 English questions generated from 458 medical case-report articles published from March through July 2026. Of these, 225 are text-only and 264 use one or more bundled image crops. In v1.2.0, the active questions were rewritten for clearer syntax, with the prior wording retained in question_text_original and a per-row audit in simplification. v1.2.1 adds one expert-reviewed target realignment for q02-fae0fd6e965a8cff: the active question now retains injury-depth interpretation instead of explicitly asking for residual-finding identification. This is a benchmark target change, not a syntax-only rewrite.

Data preview

The table below shows a deterministic sample of text-only and image-based questions. Scores are out of 10; full questions, answers, reasoning, rubrics, evidence, evaluations, correctness labels, and provenance are available in the Parquet and JSONL files.

Item	Month	Mode	Question excerpt	Type	Scores (Gemini / Opus)	Source
q01-2bf9320c0f5eeb2f	2026-03	Text	Case: In children with heterozygous loss-of-function variants in the NF-kB regulator A20 (TNFAIP3), one child had a ~2.1…	Mechanism	3 / 0	article
q01-04970952ca1605f4	2026-03	Image	Case: A 6-month-old infant presented with failure to thrive, tachypnea, cardiomegaly, and severe heart failure. Echocard…	Mechanism	4 / 6	article
q03-0e3584f359618dcf	2026-04	Text	Case: After resection, the tumor was staged pT4a because the depth of invasion was 1.2 cm with bone invasion. A multidis…	Prognosis	6 / 6	article
q01-95f1240384e9ef58	2026-04	Image	Case: A patient had a prolonged pulseless electrical activity arrest, requiring about 16 minutes of resuscitation before…	Investigation Interpretation	0 / 0	article
q02-e66d985c24a11e99	2026-05	Text	Case: A patient received myeloablative conditioning with cyclophosphamide 50 mg/kg/day for four days. Hyperhydration and…	Mechanism	5 / 5	article
q01-aadc73e1ca6da2e4	2026-05	Image	Case: A 53-year-old woman with systemic lupus erythematosus was taking chronic low-dose corticosteroids. She presented w…	Investigation Interpretation	2 / 3	article
q03-2b1b1b38967a8b31	2026-06	Text	Case: A patient taking an ALK tyrosine kinase inhibitor develops symptoms during the weeks after treatment begins. She h…	Treatment	5 / 5	article
q01-92fa6db30219ebc9	2026-06	Image	Case: A woman in her early 60s presented with flank pain, dysuria, and macroscopic haematuria. Her serum creatinine had …	Investigation Interpretation	6 / 6	article
q03-4324517b48a811a6	2026-07	Text	Case: An early CT showed a pediatric intra-abdominal cystic lesion with a multivesicular internal pattern. Hydatid disea…	Complication	4 / 4	article
q01-96a12edab9afec41	2026-07	Image	Case: A woman presented with acute chest pain. Within one hour, high-sensitivity cardiac troponin I rose from about 160 …	Investigation Interpretation	2 / 2	article

Selection

The release has two inclusion paths:

1. Primary dual-model failures (470 rows): Gemini 3.1 Pro Preview and Claude Opus 5 both scored below 7/10 in the closed-book setting.
2. Linked Q1 context (19 rows): an available Q1 was included when a Q2 or Q3 from the same source article was a primary dual-model failure. These 19 Q1 rows scored at least 7/10 for Opus and would have been omitted by the primary rule alone.

The exact frozen rule is:

gemini_closed_book_score < 7 AND opus_closed_book_score < 7

Every row has explicit correctness.gemini_closed_book.result and correctness.opus_closed_book.result fields with the values right or wrong, plus the underlying score, threshold, and boolean passed value. In this release, Gemini is labeled wrong on all 489 rows; Opus is labeled wrong on the 470 primary rows and right on the 19 linked Q1 rows.

“Right” and “wrong” are threshold labels: a score of at least 7/10 is right and a score below 7/10 is wrong. They do not assert that every medical statement in a response was true or false; a response may lose points for missing rubric requirements. The March source retains the Opus answer and detailed grading. The April-July Opus artifact contains only its scalar score, not the raw answer or judge rationale.

The Gemini and Opus scores were produced using the original v1.1.0 question wording. They were not rerun on the revised questions and should not be interpreted as performance measurements for v1.2.x prompts. For q02-fae0fd6e965a8cff, the stored 0/10 remains a v1.1 two-target result, not a measurement on the v1.2.1 one-target question.

Contents

• data/test-00000-of-00001.parquet: viewer-ready test split using Hugging Face native image features.
• data/challenge_2026_03_07.jsonl: the same records in nested JSON Lines form.
• assets/: image crops required by multimodal questions.
• metadata/selection_manifest.json: selection rule, exact models, source snapshot metadata, per-item scores, paths, and hashes.
• metadata/source_catalog.json: DOI, citation, license, official article URL, and direct PDF URL when exposed by authoritative registries or publisher metadata.
• metadata/checksums.sha256: SHA-256 manifest for the release.

The original article PDFs are intentionally not redistributed. Each record retains the original PDF hash and links to the DOI, official article page, and direct PDF where available.

Hugging Face data structure

• images: List[Image] with image bytes embedded in Parquet for direct Dataset Viewer rendering.
• input_assets: List[string] preserving the original repository-relative paths for provenance and direct downloads.
• The dataset uses the default config and test split.

from datasets import load_dataset

dataset = load_dataset(
    "txz32102/pediamed-mdle-challenge",
    split="test",
)
images = dataset[0]["images"]

Important fields

Each row retains the question, golden answer, reasoning, rubric, evidence quotes and page numbers, taxonomy, source identity and hashes, Gemini evaluations, Opus recheck, correctness labels, selection provenance, and portable asset paths. question_text contains the simplified prompt, question_text_original preserves the v1.1.0 wording, and simplification records the per-item rewrite audit. selection.label distinguishes dual_model_failure from linked_q1_context, while selection.linked_dual_model_failure_item_keys connects each companion Q1 to the Q2/Q3 failures that caused its inclusion. Machine-specific absolute paths from the source export were removed and listed in the release manifest.

Dataset caveats

• This is a deliberately difficult, model-selected set and is not representative of routine clinical questions.
• The questions are case-report-derived benchmark items, not questions from a medical school, board, or licensing examination.
• Gemini 3.1 Pro Preview served as both the original solver/judge and the Opus judge, so results may contain judge-specific selection bias.
• Source article licenses vary. Article rights remain with their respective authors and publishers. No license to redistribute the linked articles is granted here.
• This benchmark is for research evaluation only and must not be used for patient-care decisions.

Version

v1.2.1 — expert-review correction for q02-fae0fd6e965a8cff: removes the explicit residual-finding ask, retains the injury-depth target, and replaces the active rubric with a 10-point single-target rubric. Historical evaluations remain unchanged.

v1.2.0 — rewrites all 489 active questions for clearer presentation while preserving the v1.1.0 wording and per-item rewrite audit.

Citation

Please cite the accompanying PediaMed / Medical Doctor's Last Exam paper once its final PDF is available. Until then, cite this dataset by its repository URL and version.