license: cc-by-4.0
language:
- en
task_categories:
- question-answering
tags:
- clinical
- clinical-guidelines
- citation
- synthetic
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: test
path: questions.jsonl
Synthetic clinical question set for verbatim-citation evaluation
222 synthetic clinical questions, each written from one section of a published cardiometabolic practice guideline.
Questions carry the identifier of their source section, not its text. The guidelines are copyrighted and cannot be redistributed, so you need your own copy of them. The harness rebuilds the section tree from them, and the labels resolve against that tree.
Paper: arXiv:2609.15964 Code and harness: github.com/oberst-lab/verifiable-by-construction
Fields
| Field | Notes |
|---|---|
question_id |
unique, shaped <guideline_id>_<node_id>_<style>_qNN |
question |
the question text |
question_style |
lookup (asks for a recommendation directly), scenario (describes one patient and asks what applies to them), or sdm (asks how to discuss the options with a patient). 74 each |
guideline_id |
blood-pressure-2025 (42), cvd-prevention-2019 (75), diabetes-care-2026 (75), dyslipidemia-2026 (30) |
guideline_name |
the guideline's full title |
source |
the gold label: which section the question came from |
Inside source:
| Key | Notes |
|---|---|
node_id |
the gold section. Unique only together with guideline_id, since section ids repeat across guidelines |
chapter_id, subtopic_id |
the same section at chapter and at subtopic granularity |
path |
node ids from the document root down to the section |
title_path |
the corresponding section headings |
level, section_type |
position in the section tree |
char_count |
length of the section text |
74 sections in total, three questions on each. Generated with gpt-4o-mini, one
section at a time.
License
CC BY 4.0 covers the questions and labels here. The guidelines themselves belong to the bodies that issued them.
Citation
Jiashuo Zhang, Yuling Chen, Yvonne Commodore-Mensah, Michael Oberst. Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering. arXiv:2609.15964, 2026.