Hugging Face's logo

Datasets:
Filo-White
/
ITAMed

Tasks:
Question Answering
Multiple Choice
Modalities:
Tabular
Text
Formats:
json
Languages:
Italian
English
Size:
1K - 10K
Tags:
medical
clinical
exam
italian
bilingual
multimodal
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ITAMed / README.md
Filo-White's picture
Filo-White
Update README.md
8bf516e verified
preview code
|
Raw
History Blame Contribute Delete
9.41 kB
metadata 
language:
  - it
  - en
license: cc-by-4.0
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
  - multiple-choice
tags:
  - medical
  - clinical
  - exam
  - italian
  - bilingual
  - multimodal
  - mcq
  - benchmark
  - ssm
  - specialization
pretty_name: ITAMed - Italian Medical Specialization Exam Dataset
dataset_info:
  features:
    - name: year
      dtype: int32
    - name: question_number
      dtype: int32
    - name: question_code
      dtype: string
    - name: question
      dtype: string
    - name: answer_a
      dtype: string
    - name: answer_b
      dtype: string
    - name: answer_c
      dtype: string
    - name: answer_d
      dtype: string
    - name: answer_e
      dtype: string
    - name: correct_answer
      dtype: string
    - name: category
      dtype: string
    - name: question_type
      dtype: string
    - name: has_image
      dtype: bool
    - name: image_category
      dtype: string
    - name: image_path
      dtype: string
  splits:
    - name: train
      num_examples: 1260
  config_name: default
configs:
  - config_name: it
    data_files: data/ITAMed_IT.json
    default: true
  - config_name: en
    data_files: data/ITAMed_EN.json

ITAMed: Italian Medical Specialization Exam Dataset (2017–2025)

GitHub License: CC BY 4.0

Dataset Description

ITAMed is a comprehensive, bilingual dataset of 1,260 multiple-choice medical questions from the Italian National Medical Specialization Entrance Exam (Concorso SSM — Scuole di Specializzazione in Medicina), spanning 9 consecutive years (2017–2025).

Every question has been classified into one of 28 medical specialties through a rigorous dual-annotator LLM protocol validated by independent expert adjudication, and professionally translated from Italian to English with expert review.

Key Features

  • 1,260 questions — 140 per year across 9 years (2017–2025)
  • Bilingual — Original Italian + expert-reviewed English translation
  • 28 medical specialties — Dual-LLM classification (κ=0.895) with expert adjudication
  • Question-type annotation — 928 case-based (73.7%) and 332 knowledge-based (26.3%), classified by two medical specialists
  • 76 image-bearing questions — With 21 diagnostic image-type categories
  • Ready for benchmarking — Standard MCQ format with shuffled-answer utility for unbiased evaluation

Specialty Distribution

Heatmap: Questions by Specialization and Year

Combined Distribution Chart

Quick Start 

from datasets import load_dataset

# Load Italian version (default)
dataset = load_dataset("Filo-White/ITAMed", "it")

# Load English version
dataset_en = load_dataset("Filo-White/ITAMed", "en")

# Filter by specialty
cardiology = dataset["train"].filter(lambda x: "Cardiology" in x["category"])

# Filter by year
year_2024 = dataset["train"].filter(lambda x: x["year"] == 2024)

# Image-bearing questions only
with_images = dataset["train"].filter(lambda x: x["has_image"])

Available Files

Path Format Description
data/ITAMed_IT.json JSON Complete Italian dataset (1,260 questions)
data/ITAMed_EN.json JSON Complete English dataset (1,260 questions)
images/{year}/ PNG/JPG Diagnostic images referenced by questions (76 total)

The canonical repository with per-year JSON/XLSX files, distribution workbooks, and full documentation is hosted on GitHub.

Dataset Schema

Field Type Description
year int Exam year (2017–2025)
question_number int Position within the year (1–140)
question_code string Official unique question identifier
question string Full question text (clinical vignette + stem)
answer_a string Answer option A (always correct)
answer_b – answer_e string Distractor options
correct_answer string Always "A"
category string Medical specialty (1–2, semicolon-separated)
question_type string case-based or knowledge-based
has_image bool Whether the question references an image
image_category string Type of diagnostic image (21 categories)
image_path string Relative path to image file

Statistics

Metric Value
Total questions 1,260
Years covered 2017–2025 (9 years)
Questions per year 140
Case-based questions 928 (73.7%)
Knowledge-based questions 332 (26.3%)
Single-specialty questions 824 (65.4%)
Multi-specialty questions 436 (34.6%)
Questions with images 76 (6.0%)
Medical specialties 28
Image categories 21
Languages Italian + English

Top 10 Specialties (by question count)

# Specialty Count
1 Cardiology and Cardiac Surgery 154
2 Oncology 112
3 Pulmonology and Thoracic Surgery 103
4 Neurology and Neurosurgery 100
5 Pediatrics 95
6 Gastroenterology 93
7 Gynecology and Obstetrics 87
8 Infectious Diseases 84
9 General Surgery 80
10 Endocrinology 76

Yearly Image Distribution

Year 2017 2018 2019 2020 2021 2022 2023 2024 2025
Images 14 9 5 6 6 4 11 8 13

Construction Pipeline 

PDF Sources (MUR)  →  Text Extraction  →  Classification  →  Translation  →  Quality Control
   (9 PDFs)           (pdfplumber)        (dual-LLM +         (Claude +       (expert review
                                           expert review)      expert review)   + validation)
  1. PDF Extraction — Automated text extraction with pdfplumber and custom regex parsers, followed by manual verification against source documents
  2. Specialty Classification — Dual-annotator LLM protocol:
    • Independent classification by Claude claude-opus-4-8 and GPT-5.5 (Cohen's κ = 0.895, "almost perfect")
    • 380 disagreements reviewed by two independent medical specialists (inter-reviewer κ = 0.856)
    • 126 remaining discordances resolved through expert consensus
  3. Question-Type Annotation — 928 case-based and 332 knowledge-based, jointly classified by two medical specialists following operational criteria distinguishing clinical reasoning from factual recall
  4. Image Annotation — Manual identification and categorization of 76 image-bearing questions into 21 diagnostic image types
  5. Translation — Medical translation (Claude claude-opus-4-8, IT→EN) with USMLE-style English terminology
  6. Quality Control — Expert medical review of all 1,260 translations: 571 corrections across 431 questions (34.2% correction rate)

Full methodology: GitHub Repository

Data Source

Questions were extracted from the official PDF documents of the Italian National Medical Specialization Entrance Exam (Concorso per l'ammissione alle Scuole di Specializzazione in Medicina e Chirurgia), administered annually by the Italian Ministry of University and Research (MUR).

  • 2017–2019: Scenario-based format (questions grouped under shared clinical scenarios)
  • 2020–2025: Standalone format (self-contained questions)
  • Correct answer: Always option A (official release format)

Use Cases

  • LLM Medical Benchmarking — Evaluate language models on real clinical exam questions in Italian and/or English
  • Cross-lingual Medical NLP — Compare model performance across Italian and English on identical clinical content
  • Medical Education Research — Analyze question difficulty, topic distribution, and temporal trends
  • Multimodal Medical AI — Evaluate vision-language models on image-bearing clinical questions

Limitations

  • The correct answer is always in position A. A shuffle_answers.py utility (seed-based, deterministic) is provided for unbiased evaluation.
  • Image-bearing questions (6%) require the associated image for complete understanding.
  • Translation was LLM-generated and expert-reviewed; some nuances may differ from native human medical translation.

License

CC BY 4.0