README.md

metadata

pretty_name: QFR-RAG
task_categories:
  - text-retrieval
  - text-ranking
  - question-answering
language:
  - en
size_categories:
  - n<1K
tags:
  - retrieval-augmented-generation
  - adversarial-evaluation
  - abstention
  - citation-grounding
  - evidence-retrieval
  - clinical-documentation
  - technical-documentation
configs:
  - config_name: taskA
    data_files:
      - split: test
        path: taskA.jsonl
  - config_name: taskB
    data_files:
      - split: test
        path: taskB.jsonl
  - config_name: taskA_adversarial
    data_files:
      - split: test
        path: taskA_adversarial.jsonl
  - config_name: taskB_adversarial
    data_files:
      - split: test
        path: taskB_adversarial.jsonl

QFR-RAG

QFR-RAG is a manually curated evaluation resource for retrieval-augmented generation in specialised technical-medical documentation. The dataset focuses on Quantitative Flow Ratio (QFR), where answering user questions often requires precise evidence from software documentation, angiographic acquisition requirements, methodological material, and related technical-clinical sources.

The resource is designed to evaluate whether RAG systems can retrieve sufficient evidence, generate grounded answers, cover required answer content, cite relevant support, abstain when evidence is insufficient, and handle adversarial questions safely.

The released dataset contains questions, reference answers, evidence annotations, answer-nugget annotations, and adversarial evaluation metadata. It does not redistribute proprietary or restricted source documentation. Users are responsible for obtaining and using any source documents in accordance with their original licenses and access conditions.

Dataset Structure

QFR-RAG contains three components: Task A, Task B and adversarial questions.

Task A: Technical Extraction

Task A contains 50 questions targeting specific technical facts from QFR-related documentation. These questions typically ask for precise parameters, definitions, procedural requirements or software-specific constraints.

Task A is intended for evaluating exact evidence retrieval and documented answer correctness.

Typical fields include:

• id
• task
• question
• reference_answer
• evidence_chunk_ids

Task B: Multi-Evidence Clinical/Technical QA

Task B contains 50 questions that require combining information from multiple pieces of evidence. The questions are phrased as realistic information needs, for example from a user of QFR software or someone learning about QFR methodology.

Task B contains 103 required evidence slots and 278 atomic answer nuggets. Evidence slots are used to evaluate whether retrieval covers the required evidence, while answer nuggets are used to evaluate whether generated answers contain the required information.

Typical fields include:

• id
• task
• question
• reference_answer
• evidence_slots
• answer_nuggets
• evidence_chunk_ids

Adversarial Questions

The adversarial component contains 300 paired questions derived from the base Task A and Task B questions. These examples are designed to test whether systems can avoid unsafe or unsupported behaviour when a question should not simply be answered as stated.

The adversarial questions cover three categories:

• Nonsensical questions: questions that are incoherent, impossible or not meaningful in the QFR context. A good system should not invent an answer.
• False-premise questions: questions that assume an incorrect or unsupported claim. A good system should correct or reject the premise instead of accepting it.
• Safety-critical questions: questions that could lead to unsafe technical or clinical guidance if answered carelessly. A good system should avoid unsupported recommendations and should stay grounded in the available evidence.

The public adversarial files include:

• id
• base_id
• task
• base_task
• adversarial_category
• question
• target_response_type

Intended Use

QFR-RAG is intended for research on evidence retrieval, reranking, grounded question answering, citation grounding, answer completeness, abstention and adversarial robustness in technical-medical RAG settings.

Suitable uses include:

• evaluating retrievers and rerankers;
• evaluating RAG answer generation;
• comparing standard RAG and oracle-evidence settings;
• measuring evidence sufficiency and answer completeness;
• testing whether systems abstain when evidence is missing;
• testing robustness to false-premise, nonsensical, and safety-critical prompts.

The dataset is intended as a diagnostic benchmark, not as a general medical QA dataset.

Out-of-Scope Use

QFR-RAG should not be used as a source of medical advice, clinical decision support, or procedural guidance. It should not be used to train or deploy systems that make patient-specific recommendations or replace expert clinical judgement.

The dataset also should not be treated as a complete representation of QFR, coronary physiology, or QFR software use. It reflects a fixed curated documentation corpus and a manually designed set of evaluation questions.

Source Data

The dataset was created from a curated collection of QFR-related technical-medical documents, including software documentation, release notes, educational material, angiography training material and QFR methodological literature.

The source documents were segmented into section-aware passages with stable passage identifiers.

The full source documents are not redistributed in this dataset. The released dataset contains evaluation questions, reference answers, annotations, evidence identifiers and adversarial metadata. Users are responsible for obtaining and using any source documents in accordance with their original licenses and access conditions.

Document Sources

Source	Category	Role in corpus
QFR 3.0 User Manual	Software documentation	Operational reference for QFR workflow, software use, analysis steps, and technical constraints.
QFR release notes	Software documentation	Version-specific software notes and changes.
QFR 3.0 educational module	Educational material	Supporting explanation of QFR concepts and workflow.
Basic coronary angiography training material	Educational material	Background information on coronary angiography and image acquisition.
Angiographic projections training material	Educational material	Background information on projection selection and angiographic interpretation.
Tu et al. (2012)	Methodological literature	Validation material related to three-dimensional quantitative coronary angiography methodology.
Tu et al. (2014)	Methodological literature	Evidence on fractional flow reserve calculation from three-dimensional quantitative coronary angiography and TIMI frame count.
Lansky and Pietras (2014)	Clinical commentary	Clinical context for image-based functional assessment.
Tu et al. (2015)	Review literature	Broader methodological and clinical context for image-based fractional flow reserve.
Tu et al. (2015)	Methodological literature	Evidence related to bifurcation anatomy and physiological modelling.

Annotations

The dataset includes manually created annotations.

Task A uses chunk-level evidence annotations to identify the documentation passages needed to answer each technical extraction question.

Task B uses two levels of annotation:

• Evidence slots: required evidence components for answering a question sufficiently.
• Answer nuggets: atomic answer units used to evaluate answer completeness.

The adversarial set includes category labels and target response types that describe the expected behaviour, such as answering from evidence, correcting a false premise, abstaining or refusing unsafe unsupported guidance.

Personal and Sensitive Information

The dataset does not contain patient records, patient-level clinical data, personal health information or private user data. It is based on technical-medical documentation and manually written evaluation questions.

The dataset concerns a clinical technology domain and may include medically relevant terminology. It should therefore be used carefully and only for research and evaluation purposes.

License

The released dataset files are shared under the Creative Commons Attribution-NonCommercial 4.0 International license.

This license applies only to the released QFR-RAG dataset files, such as the questions, annotations, labels, and metadata. It does not apply to the original source documents, software manuals, publications, or other third-party materials from which the evaluation resource was curated.

Source References

The following references identify the main documents and publications used to construct the QFR-RAG corpus. They are listed for transparency. The full source documents are not redistributed as part of this dataset.

• Medis Medical Imaging Systems bv. (2024). QFR 3.0 User Manual. Version QFR 3.0.26.4. Accessed 2026-06-03. https://medisimaging.com/wp-content/uploads/2024/07/QFRUserManual_3.0.26.4.pdf

• Medis Medical Imaging Systems bv. QFR release notes. Software documentation.

• Medis Medical Imaging Systems bv. QFR 3.0 educational module. Educational material.

• Shavelle, D. (n.d.). Basic Coronary Angiography: All Slides. Educational slide deck accessed via Scribd; original source and license not independently verified. Accessed 2026-06-03. https://www.scribd.com/document/386807683/Basic-Coronary-Angiography-All-Slides-pdf

• Kern, M. J. (2011). Angiographic Projections Made Simple: An Easy Guide to Understanding Oblique Views. Cath Lab Digest. Accessed 2026-06-03. https://www.hmpgloballearningnetwork.com/site/cathlab/articles/angiographic-projections-made-simple-easy-guide-understanding-oblique-views

• Tu, S., Xu, L., Ligthart, J., Xu, B., Witberg, K., Sun, Z., Koning, G., Reiber, J. H. C., & Regar, E. (2012). In vivo comparison of arterial lumen dimensions assessed by co-registered three-dimensional (3D) quantitative coronary angiography, intravascular ultrasound and optical coherence tomography. International Journal of Cardiac Imaging, 28(6), 1315–1327. https://doi.org/10.1007/s10554-012-0016-6

• Tu, S., Barbato, E., Köszegi, Z., Yang, J., Sun, Z., Holm, N. R., Tar, B., Li, Y., Rusinaru, D., Wijns, W., & Reiber, J. H. C. (2014). Fractional flow reserve calculation from 3-dimensional quantitative coronary angiography and TIMI frame count. JACC: Cardiovascular Interventions, 7(7), 768–777. https://doi.org/10.1016/j.jcin.2014.03.004

• Lansky, A. J., & Pietras, C. (2014). Fractional flow reserve from 3-dimensional quantitative coronary angiography. JACC: Cardiovascular Interventions, 7(7), 778–780. https://doi.org/10.1016/j.jcin.2014.05.002

• Tu, S., Bourantas, C. V., Nørgaard, B. L., Kassab, G. S., Koo, B.-K., & Reiber, J. H. C. (2015). Image-based assessment of fractional flow reserve. EuroIntervention, 11(V), V50–V54. https://doi.org/10.4244/eijv11sva11

• Tu, S., Echavarria-Pinto, M., von Birgelen, C., Holm, N. R., Pyxaras, S. A., Kumsars, I., Lam, M. K., Valkenburg, I., Toth, G. G., Li, Y., Escaned, J., Wijns, W., & Reiber, J. H. C. (2015). Fractional flow reserve and coronary bifurcation anatomy. JACC: Cardiovascular Interventions, 8(4), 564–574. https://doi.org/10.1016/j.jcin.2014.12.232

Citation

A citation for QFR-RAG will be added if and when an associated paper or dataset publication becomes publicly available.