room-b007/test-medicina

Medschool-Test, or "Test di Medicina"

Is your LLM able to pass a National Entrance Exam for the Italian Medical School?

This the GitHub repo for our Hugging Face dataset designed for evaluating Large Language Models (LLMs) on a broad range of questions from the national entrance exams for the Italian medical school (ORIGINAL WEBSITE). The dataset includes multiple-choice questions from various subjects such as biology, chemistry, physics, mathematics, world knowledge, and more. Each question is accompanied by five answer choices, with one correct answer. The following is an example:

{
  "id": 1691,
  "topic": "biologia",
  "text": "Come sono definite le cellule staminali che sono in grado di differenziarsi in tutti i tipi di cellule presenti nel corpo umano, ma non possono dare origine ad un organismo completo?",
  "answers": [
    "Cellule Staminali Multipotenti",
    "Cellule Staminali Pluripotenti",
    "Cellule Staminali Totipotenti",
    "Cellule Staminali Unipotenti",
    "Cellule Staminali Oligopotenti"
  ],
  "label": 1
}

The dataset is available on Hugging Face 🤗! 👉👉 LINK 👈👈.

Features

• Italian (and English coming soon): The dataset is currently available in Italian, with an English version coming soon.

• Multiple topics: Questions cover a wide range of subjects, including biology, chemistry, physics, mathematics, world knowledge (with a focus on Italian culture), and more.

• "Multiple-choice" and "Cloze-style" formats: Each question has five answer choices, with one correct answer. This dataset is designed to evaluate LLMs on both multiple-choice and cloze-style questions. More specifically, each question is presented in the following formats:

  • Multiple-choice: The question is followed by five answer choices, with one correct answer. The evaluation metrics is based on the model's ability to select the correct answer when presented with the question and answer choices.

  • Cloze-style: The question is NOT followed by answer choices. The evaluation metric is based on the model's ability to generate the correct answer when presented with the question only.

• Large-scale: The dataset contains over 3K high-quality questions, making it suitable for the evaluation of LLMs.

Evaluation

The dataset is designed to evaluate LLMs on a wide range of questions from medical school entrance exams. The evaluation metrics are based on the model's ability to select the correct answer when presented with the question and answer choices (multiple-choice format) or generate the correct answer when presented with the question only (cloze-style format).

Leaderboard

Coming soon!

Scoring

For each question, the model obtains a score ranging from -0.4 to 1.5, based on the following criteria:

• Correct Answer: If the model selects the correct answer, it receives a score of 1.5.
• Incorrect Answer: If the model selects an incorrect answer, it receives a score of -0.4.
• No Answer: If the model does not select any answer, it receives a score of 0.0.

The final score is calculated as the weighted average of the scores obtained on all questions, where the weight of each question depends on the topic of the question, as follows:

Subject	Score	Weight
Biology	23/60	0.3833
Chemistry	15/60	0.25
Math & Physics	13/60	0.2167
World Knowledge	4/60	0.0667
Logic & Reasoning	5/60	0.0833

This is the same scoring system used for the official Italian medical school entrance exams.

Evaluation Script

The evaluation script is available on our GitHub repo! Check it out!

The evaluation is based on the lm-evaluation-harness library, which provides a simple and flexible way to evaluate LLMs on a wide range of tasks and datasets. The tasks are defined in tasks/medschool-test.

Requirements and Installation

We recommend using Conda to create a new environment and install the required libraries. You can create a new Conda environment using the following command:

conda create -n medschool-test python=3.10
conda activate medschool-test

Using Conda is optional but highly recommended to avoid conflicts with existing libraries.

To run the evaluation, you need to install the lm-evaluation-harness library and the transformers library. You can install them using the following command:

pip install --upgrade -r requirements.txt

Running the Evaluation Script

To run the evaluation, you can use the following command:

# Model to evaluate: meta-llama/Meta-Llama-3.1-8B-Instruct (in bfloat16)
MODEL_ARGS="pretrained=meta-llama/Meta-Llama-3.1-8B-Instruct,dtype=bfloat16"

# Evaluate the model on both multiple-choice and cloze-style formats
TASKS="medschool_test_it_mc,medschool_test_it_cloze"

# Create the output directory if it does not exist
OUTPUT_DIR="outputs/"
mkdir -p $OUTPUT_DIR

# Run the evaluation with the lm-evaluation-harness library
accelerate launch -m lm_eval \
  --model hf \
  --model_args $MODEL_ARGS \
  --tasks $TASKS \
  --batch_size auto \
  --log_samples \
  --output_path $OUTPUT_DIR \
  --include tasks/medschool-test/

This command evaluates the model meta-llama/Meta-Llama-3.1-8B-Instruct on the Italian version of the dataset in both multiple-choice and cloze-style formats. The evaluation results are saved in the outputs/ directory.

Please, refer to the examples in examples/evaluation or lm-evaluation-harness repository for more details on how to use the library.

The scores provided by the evaluation script are 5, one for each subject, and they are between -0.4 and 1.5. They, the subject-scores, are computed as the average score over all the questions for each specific subject. To get the final score, you need to calculate the weighted average of the subject-scores, where the weight of each subject-score depends on the subject itself (as described in the "Scoring" section above).

For example, if the model obtains the following subject-scores:

• Biology: 1.0571
• Chemistry: 0.7598
• Knowledge: 1.0518
• Reasoning: 0.2005
• Math & Physics: 0.4302

The final score is calculated as follows:

average_score_per_question = (1.0571 * 0.3833) + (0.7598 * 0.25) + (1.0518 * 0.0667) + (0.2005 * 0.0833) + (0.4302 * 0.2167) = 0.7752
overall_score = average_per_question * 60 = 46.51

The maximum possible score is 90 = 60 * 1.5 , when all the answers are correct, while the minum score is -24 = 60 * -0.4 , when all the answers are incorrect.

Data

Source

The dataset is collected from the official Italian website of the Ministry of Education, University and Research (MIUR), which hosts a large collection of past entrance exams for medical school in Italy. The dataset includes questions from various subjects, such as biology, chemistry, physics, mathematics, world knowledge, and more. You can find the original dataset HERE.

Composition

The dataset contains over 3K questions, making it suitable for training and evaluating LLMs. The following table shows the number of questions in the dataset for each subject:

Subject	Number of questions
Biology	1,180
Chemistry	1,009
Math & Physics	655
World knowledge	245
Logic & Reasoning	212
Total	3,301

Size comparison with other benchmarks

The dataset contains over 3K questions, making it suitable for training and evaluating LLMs. The following table shows the size of the dataset compared to other benchmarks:

Dataset	Number of questions
Medical School Entrance Exams	~3,300
ARC-Challenge (test)	~1,170
ARC-Easy (test)	~2,380
BoolQ (validation)	~3,270
CommonSenseQA (validation)	~1,220
GSM8K (test)	~1,320
MMLU (test)	~14,000
PIQA (validation)	~1,850
SciQ (test)	~1,000
TruthfulQA (test)	~820
WinoGrande (test)	~1,767

Format

The dataset is provided in JSONL format, with each line representing a single question in the following format:

{
  "id": 1691,
  "topic": "biologia",
  "text": "Come sono definite le cellule staminali che sono in grado di differenziarsi in tutti i tipi di cellule presenti nel corpo umano, ma non possono dare origine ad un organismo completo?",
  "answers": [
    "Cellule Staminali Multipotenti",
    "Cellule Staminali Pluripotenti",
    "Cellule Staminali Totipotenti",
    "Cellule Staminali Unipotenti",
    "Cellule Staminali Oligopotenti"
  ],
  "label": 1
}

where:

• id (int): The unique identifier of the question.
• topic (str): The subject of the question.
• text (str): The text of the question.
• answers (list): A list of five answer choices, with one correct answer.
• label (int): The index of the correct answer in the answers list (0-indexed).

Examples

Biology

{
  "id": 1691,
  "topic": "biologia",
  "text": "Come sono definite le cellule staminali che sono in grado di differenziarsi in tutti i tipi di cellule presenti nel corpo umano, ma non possono dare origine ad un organismo completo?",
  "answers": [
    "Cellule Staminali Multipotenti",
    "Cellule Staminali Pluripotenti",
    "Cellule Staminali Totipotenti",
    "Cellule Staminali Unipotenti",
    "Cellule Staminali Oligopotenti"
  ],
  "label": 1
}

Chemistry

{
  "id": 19,
  "topic": "chimica",
  "text": "Rispetto alla classificazione che si trova nella tavola periodica il fluoro fa parte del: ",
  "answers": [
    "gruppo dei gas nobili",
    "gruppo degli alogeni",
    "gruppo dei lantanidi",
    "secondo gruppo",
    "quarto periodo"
  ],
  "label": 1
}

Math & Physics

{
  "id": 1203,
  "topic": "fisica-matematica",
  "text": "Quali sono le coordinate del centro della circonferenza di equazione x^2 + y^2 + 2x – 6y + 5 = 0?",
  "answers": [
    "(2 ; –6)",
    "(–2 ; 6)",
    "(1 ; 3)",
    "(–1 ; 3)",
    "(2 ; 3)"
  ],
  "label": 3
}

World Knowledge

{
  "id": 11,
  "topic": "competenze-conoscenze",
  "text": "Quale dei seguenti è il primo romanzo di Italo Calvino, pubblicato nel 1947?",
  "answers": [
    "Se una notte d'inverno un viaggiatore",
    "Il barone rampante",
    "Palomar",
    "Le cosmicomiche",
    "Il sentiero dei nidi di ragno"
  ],
  "label": 4
}

Logic & Reasoning

{
  "id": 2539,
  "topic": "logica",
  "text": "Tutti i pasticcieri praticano il kendo; Gianluca pratica il kendo. Quale delle seguenti affermazioni aggiuntive consentirebbe di dedurre con certezza che Gianluca è un pasticciere?",
  "answers": [
    "Tra le persone che praticano kendo vi sono dei pasticcieri",
    "Alcune persone che praticano kendo si chiamano Gianluca",
    "Alcune persone che praticano kendo sono pasticcieri",
    "Non è certo che ogni persona che pratica kendo sia anche un pasticciere",
    "Ogni persona che pratica kendo è anche un pasticciere"
  ],
  "label": 4
}

Reproducibility

We provide the code to reproduce our dataset in the src/data/collection directory. The code is written in Python and uses the beautifulsoup4 library to scrape the questions from the official MIUR website. You can run the code to collect the latest questions from the website and generate the dataset in JSONL format.

You can run the code using the following command:

python src/data/collection/collect_questions.py \
    --output_path data/questions/medical_school_questions.jsonl \
    --last_page_index 174

This command collects the questions from the official MIUR website and saves them in the data/questions/medical_school_questions.jsonl file. You can specify the number of pages to scrape using the --last_page_index argument (where each page contains 20 questions and the last page is 174).

Citation

We are currently writing a report on the dataset and will provide the citation information soon.

License

This dataset is released under the Apache 2.0 License. Feel free to use it for research, commercial, or personal purposes.