Datasets:

Modalities:
Text
Formats:
json
Languages:
Italian
Libraries:
Datasets
pandas
MED-ITA / README.md
DanielePoterti's picture
Fix dataset viewer configuration
b48f4c8 verified
metadata
dataset_info:
  features:
    - name: question
      dtype: string
    - name: options
      struct:
        - name: A
          dtype: string
        - name: B
          dtype: string
        - name: C
          dtype: string
        - name: D
          dtype: string
        - name: E
          dtype: string
    - name: answer
      dtype: string
    - name: category
      dtype: string
    - name: macrocategory
      dtype: string
  splits:
    - name: train
      num_bytes: 2774754
      num_examples: 10000
  download_size: 1655237
  dataset_size: 2774754
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/dataset.jsonl
language:
  - it

Dataset Card for MED-ITA

We introduce MED-ITA, a large-scale evaluation benchmark designed to assess the medical knowledge and understanding capabilities of Large Language Models (LLMs) in the Italian language.

Above are example questions from MED-ITA. Note: every example is a direct translation; the original questions are in Italian. The correct option is marked by (✓).

Dataset Details

Dataset Description

The benchmark comprises 10,000 multiple-choice questions drawn from official national examinations used to certify and recruit medical professionals across the Italian healthcare system. MED-ITA covers 50 distinct clinical and biomedical categories, ranging from core medical sciences (e.g., pathological anatomy, pharmacology and clinical toxicology, clinical pathology) to specialised clinical disciplines (e.g., neurology, cardiac surgery, psychiatry) and cross disciplinary fields (e.g., legal medicine, health statistics and biometry).

MED-ITA offers the first fine-grained, domain-specific benchmark for Italian medical language understanding, providing a foundation for advancing culturally and clinically grounded AI applications in healthcare.

How to Use

First, clone the MED-ITA repository to your local machine:

git clone https://github.com/Crisp-Unimib/MED-ITA
cd MED-ITA

To avoid conflicts with system packages, it's recommended to use a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows use: venv\Scripts\activate

MED-ITA requires some packages to work. Install them with:

pip install -r requirements.txt

MED-ITA requires vLLM for serving local models. Install it with:

pip install vllm[all]

Run the vLLM server with your preferred model. Here’s an example using meta-llama/Llama-3.3-70B-Instruct:

vllm serve meta-llama/Llama-3.3-70B-Instruct

Once the vLLM server is running, you can tweak setting in config.yaml, such as model name, API endpoint, temperature, max tokens. You can then execute the evaluation script:

python run_eval.py

To run the 100 hard questions (which include the top 2 hardest from each category), you need to update the configuration file to point to the appropriate dataset.

Open config.yaml and modify the data_file field under the data section as shown below:

data:
  data_file: ./dataset_hard.jsonl # <-- Change this line
  output_dir: results

Dataset Sources

Dataset Structure

MED-ITA contains 10,000 carefully curated questions selected from an initial corpus of 104,876 questions.

Each question is formatted as a multiple-choice query, with an average question length of 67.6 characters and 5 answer options. The longest question is 700 characters long. The total number of tokens in the input data, which includes the question, options, and prompt template, is 1,719,397.

Column Data Type Description
question [String] The actual content of the question
options [Dict] The options to choose from. Only one is correct
answer [String] The correct answer out of the options
category [String] The dedicated category of the question
macro_category [String] The macro category of the question

Dataset Creation

Curation Rationale

The corpus comprises questions and tasks from real-world exams, professional assessments, and domain-specific challenges. Given that the data originates from institutional sources, it is expected to maintain a high standard of quality and accuracy, as domain experts crafted it for public evaluations.

Source Data

Data Collection and Processing

All documents were freely available and published between 2010 and 2024, typically in PDF, DOC, or HTML formats.

Please consult the full paper for a detailed description of our curation process.

Who are the source data producers?

The raw dataset was compiled from official public sources, including national and regional public recruitment competitions, admission tests for medical degree programs and speciality schools, and official exam preparation materials published by public agencies.

Personal and Sensitive Information

The dataset does not contain confidential information. It is also free from content that could be considered offensive, insulting, threatening, or distressing. Since it solely comprises data from standardised tests and does not involve human subjects or personal data, an ethical review process was not required.

Bias, Risks, and Limitations

MED-ITA is designed strictly for research evaluation purposes and should not be used to validate LLMs for direct clinical decision-making or patient care applications. The benchmark assesses knowledge recall and reasoning on standardised questions, which differ significantly from real-world clinical practice scenarios.

Maintenance

To ensure responsible use of MED-ITA, the research team commits to making the complete dataset, evaluation protocols, and supplementary materials freely available under permissive licensing; maintaining clear versioning and update protocols to ensure longitudinal research validity and encouraging collaborative improvements and error reporting from the broader research community.

Citation

BibTeX:

SUBMITTED AT KDD

APA:

SUBMITTED AT KDD

Dataset Card Contact

Andrea Seveso - andrea.seveso@unimib.it