NorMedQA

NorMedQA tests how well language models answer medical questions in Norwegian. It has two configurations:

• exam: 2,260 multiple-choice questions from the Norwegian national written final exam in medicine, 16 sittings from spring 2017 to spring 2026. Most come in both Bokmål and Nynorsk, making 3,940 rows. Every question has the official answer key and the official explanation for each option. 457 rows include an image, such as an ECG, X-ray, CT or clinical photo.
• sami_reading: 1,662 questions about health texts in North Sámi (1,024), Lule Sámi (345) and South Sámi (293). The passages are written by people; the questions and options are in Bokmål.
• sami_reading_sami: the same 1,662 questions with question and options in the passage's Sámi language.
• exam_sami: 200 exam questions from all 16 sittings in North Sámi.

The Sámi questions start as machine drafts (translation == "machine_draft", with uncertain_terms). Sámi-speaking physicians are reviewing them, and reviewed rows switch to translation == "reviewed" with the corrected text (progress). Treat Sámi scores as provisional until most rows are reviewed.

Code, sources and the build pipeline: https://github.com/kelkalot/normedqa

from datasets import load_dataset

exam = load_dataset("kelkalot/normedqa", "exam")
sami = load_dataset("kelkalot/normedqa", "sami_reading_sami", split="test")  # Sámi questions

Scoring

Score rows with status == "ok". A prediction is correct when it's any label in answer. More than one label appears when the key was changed at grading; the key before the change is kept in original_answer. 101 exam questions were voided at grading (status == "voided"). Three Nynorsk rows are excluded because the source document is missing a question sentence or an option.

The exam splits follow the sittings: train is 2017V–2024H, validation is 2025V and test is 2025H and 2026V. Both written forms of a question always land in the same split. Nine questions were reused word for word in a later sitting, two of them in the test split; also_in lists the other copy.

In sami_reading, gpt-5.4-mini answers 1,281 of the 1,662 questions correctly without seeing the passage. Those rows carry the flag answerable_without_passage. The other 381 need the passage, and each of them was checked against the original text, so use those to test reading of the Sámi text itself.

Baselines

Model	Exam, all scored rows	Exam, test split
gpt-5.4-mini	92.4%	91.7%
gpt-5.5	not run	95.1%

Both used low reasoning effort. Bokmål and Nynorsk scores are within one point of each other.

How it was checked

The exam rows were parsed directly from the published answer keys, which mark the correct option with an X; no language model was involved. They were then compared with three independent sources:

• the unkeyed copy of each exam, which matches in question and option text for all 3,940 rows
• the key of Fagprøven, which uses the same paper: 2,794 of 2,800 rows agree, and the other 6 are 3 questions Fagprøven graded differently
• the other written form of each question

A vision model also compared 172 random rows with the rendered PDF pages and found no wrong answers or images.

Sources

Exam questions. These come from Felles avsluttende skriftlig deleksamen i medisin (called Nasjonal delprøve i medisin in 2017–2020), from the answer keys published after grading. The documents are listed at https://www.uio.no/studier/program/medisin/tidligere-eksamensoppgaver/felles-avsluttende-deleksamen/. Each row links its source document (source_url) and the pages it was taken from. The original PDFs are not redistributed.

Sámi texts. These come from the open text collections of GiellaLT, UiT Norgga árktalaš universitehta, at the commits recorded in each row. The original publishers are:

• Norway: Helsenorge, Folkehelseinstituttet (FHI), NKVTS, Nasjonalforeningen for folkehelsen, Nasjonal kompetansetjeneste for aldring og helse, Sámediggi / Sametinget, Regjeringen, Sámi allaskuvla, Árran julevsáme guovdásj, Røyrvik kommune, Lokalavisa NordSalten and Sámi logut muitalit
• Sweden: 1177 Vårdguiden, Folkhälsomyndigheten, Region Jämtland Härjedalen and Länsstyrelsen
• Finland: Sámediggi / Saamelaiskäräjät, Saamelaismuseo Siida and a thesis in DiVA

source_publisher names the publisher for every row. The texts are used for non-commercial research with attribution.

The sami_reading questions were written by gpt-5.5 and gpt-5.4-mini from the original-language version of each passage. A question was kept only if its supporting quotes were found in both texts and a second model reached the same answer from the original.

How this release was made

The pipeline and checks were built with Claude Code (Claude Opus 5.5). The sami_reading questions were written by gpt-5.5 and gpt-5.4-mini and audited by Claude; the exam questions and answers come straight from the published answer keys.

Licence

CC BY-NC 4.0. Credit NorMedQA and the original sources listed above.

Citation

@dataset{normedqa_v3_2026,
  author       = {Riegler, Michael Alexander and Gautam, Sushant and Gran Erke, Maja},
  title        = {{NorMedQA: Norwegian Medical Question Answering Benchmark}},
  year         = 2026,
  publisher    = {Hugging Face},
  version      = {3.0},
  doi          = {10.57967/hf/11462},
  url          = {https://doi.org/10.57967/hf/11462}
}

v2: Riegler, M. A. and Gautam, S. (2025). Norwegian Medical Question Answering Dataset – NorMedQA. Zenodo. https://doi.org/10.5281/zenodo.15320037