CareMedEval dataset

CareMedEval dataset: Evaluating Critical Appraisal and Reasoning in the Medical Field

	Links
Repository	GitHub, HuggingFace
Paper	https://arxiv.org/abs/2511.03441
Contact	Doria BONZI, Alexandre GUIGGI

Description

CareMedEval (Critical appraisal and Reasoning Medical Evaluation) is a French Multiple Choice Question Answering (MCQA) dataset focused on evaluating critical appraisal skills in the medical field for scientific articles, as practiced in French medical education.

This dataset is composed of 534 questions taken from Lecture Critique d'Articles (LCA, "Critical Appraisal of research Articles) exams, taken by sixth-year medical students in France. During the LCA exams, students are asked to answer a series of multiple-choice questions grounded in the critical reading of a given scientific article.

Source data

This dataset is built from two main sources:

• Epreuves Classantes Nationales (ECN) website, where we can find the official national LCA exams
• Collège National des Enseignants de Thérapeutique (CNET) website, which publishes mock LCA exams reviewed and approved by the CNET's educational committee to ensure alignment with real LCA exams.

The articles used in these exams are publicly available online.

Figure. CareMedEval dataset constitution pipeline.

Languages

The scientific articles are in English. The questions, answers, and justifications are originally in French. You can find in ./english_only/ a version of this dataset with automatically translated questions, answers and justifications, from French to English. We used Gemini 2.5 Flash for this translation.

Annotations and labels

Each question in the CareMedEval dataset was manually annotated with one or more labels, reflecting the cognitive and analytical skills required to answer it. Question labeling was conducted by one annotator with a background in general medecine, using reference textbooks and their medical expertise. The labels are defined as follows:

Label	Description	Skills required
design	Identification of study design	Information retrieval
statistics	Understanding and interpretation of statistics	General knowledge, Information retrieval
methodology	Knowledge of scientific methodology	General conceptual understanding
limitations	Critical appraisal of biases and limitations	Contextual reasoning
applicability	Clinical relevance and applicability	Contextual reasoning

Since questions often target multiple dimensions of skills, they can be assigned more than one label. These labels can help to determine if certain categories of questions are more challenging than others for models, depending on the types of skills they require.

For data sourced from ECN website, we manually corrected the exam questions with the help of a general practitioner, as the official answer keys were not publicly available for these past exams.

For data from CNET website, we also collected the correction and justifications provided by medical professionnals for some questions. These justifications explain why some answers are correct or false, offering valuable insights into clinical reasoning and critical appraisal.

Dataset structure

• To access the full dataset, go to ./data/all/, which contains both /articles/ and /questions.json.
• To access only the questions that include human justifications (204 questions out of 534), go to ./data/with_justifications/.
• To access our manually annotated subset of 16 questions indicating whether questions require context (articles) to be answered correctly, go to ./data/requires_context/.

Questions

Example of a question:

        "id": "fd30e0bb8cbe1d73039ffebab8e98019836a2e1af03b383f48be020d79f61310",
        "id_article": "article_3",
        "source_exam": "https://therap.fr/lca-mai-2019/",
        "date_exam": "2019",
        "article_link": "https://jamanetwork.com/journals/jama/fullarticle/2547755",
        "article_date": "2016",
        "question": "La publication de cet essai négatif :",
        "answers": {
            "a": "Est une situation rare les essais négatifs sont le plus souvent non publiés",
            "b": "Est inutile puisque n’apporte aucune information sur le traitement en pratique",
            "c": "Doit faire cesser toute investigation ultérieure sur cette molécule dans cette indication",
            "d": "Permettra d’intégrer plus facilement cet essai dans une revue systématique/metaanalyse",
            "e": "Est logique la publication des résultats de tous les essais est obligatoire"
        },
        "correct_answers": [
            "a",
            "d"
        ],
        "essential_answers": [],
        "unacceptable_answers": [],
        "labels": [
            "limitations",
            "applicability"
        ],
        "justification": "50% des essais sont non publiés et ce surtout s'ils sont négatifs. Savoir qu’un traitement n’est pas supérieur à un autre est en soi une information importante. Voir la discussion de l’essai. Une revue systématique pour donner un reflet réel de l’ensemble de la recherche sur une question doit inclure tous les essais réalisés qu’ils soient positifs ou négatifs, sinon l’éventuelle méta-analyse donnera un résultat biaisé. Il est plus simple de trouver les résultats si l’essai est publié. Hautement souhaitable et recommandée.",
        "nb_correct_answers": 2

Each question in the dataset is represented as a JSON object with the following fields:

Field	Description
id	A unique identifier for the question
id_article	An internal ID linking the question to the corresponding article in the dataset
source_exam	The URL of the exam or online resource (e.g., CNET or ECN) from which the question was extracted
date_exam	The date on which the exam was presented
article_link	The direct link to the full-text scientific article used as the basis for the question
article_date	The date on which the article was published
question	The multiple-choice question (MCQ) as presented in the exam
answers	A dictionary mapping each option label (A to E) to its full answer text
correct_answers	A list of option labels corresponding to the correct answers. Questions can have one or multiple correct options
essential_answers	A list of essential answers. If any of these are not selected, the question is considered incorrect in LCA grading
unacceptable_answers	A list of inadmissible answers. If any of these are selected, the question is considered incorrect due to the severity of the error in LCA grading
labels	A list of strings, indicating the types of reasoning or knowledge required to answer the question (see label table above)
justification	An expert-written explanation justifying the correct (and sometimes incorrect) answers; available for CNET-sourced questions
requires_context	Boolean indicating if the question requires the article to be answered correctly or not
nb_correct_answers	The number of correct options for the question

Articles

Each question in the dataset is linked to a scientific article through the id_article field, which serves as a reference key to the corresponding article file. Articles are available in plain text (.txt) format for easier processing, and the original PDF versions are also included. In addition, each question includes a direct article_link field pointing to the online HTML version. To provide a more concise context, we extracted the abstracts of the articles and stored them in ./data/all/articles_abstract, available in plain text (.txt) format.

Dataset benchmark

Figure. Overview of the model evaluation pipeline of the CareMedEval dataset.

We conducted a series of experiments on our dataset to evaluate the critical appraisal skills of a diverse pool of LLMs. While our dataset can support multiple experimental setups, we focus here on a MCQA task for reference results, based solely on the textual content of the articles. Details on our experimental protocol are available on TBA.

Results

Results of long-context models on the biomedical MCQA task of our dataset: zero-shot with full-text article provided. Bold values indicate the best results per metric, and underlined values the second-best. Baselines include random or frequency-based answer selection strategies.

Comparison of model performance according to "reasoning" tokens generation level on the CareMedEval benchmark.

Citation

To cite this dataset:

@misc{bonzi2025caremedevaldatasetevaluatingcritical,
      title={CareMedEval dataset: Evaluating Critical Appraisal and Reasoning in the Biomedical Field}, 
      author={Doria Bonzi and Alexandre Guiggi and Frédéric Béchet and Carlos Ramisch and Benoit Favre},
      year={2025},
      eprint={2511.03441},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.03441}, 
}