Update README.md

README.md

@@ -28,11 +28,30 @@ pretty_name: Test di Medicina
 
 ## Is your LLM able to pass a National Entrance Exam for the Italian Medical School?
 
-This a Huggingface dataset designed for evaluating Language Model (LLM) on a broad range of questions from the **national entrance exams 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](https://domande-ap.mur.gov.it/domande)).
 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.
+Each question is accompanied by five answer choices, with one correct answer. The following is an example:
+
+```json
+{
+  "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](https://huggingface.co/datasets/room-b007/test-medicina) 👈👈.
 
 ## 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.
 
@@ -44,9 +63,6 @@ Each question is accompanied by five answer choices, with one correct answer.
 
 - **Large-scale**: The dataset contains over 3K high-quality questions, making it suitable for the evaluation of LLMs.
 
-- **Italian (and English coming soon)**: The dataset is currently available in Italian, with an English version coming soon.
-
-
 ## 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).
@@ -75,37 +91,81 @@ This is the same scoring system used for the official Italian medical school ent
 
 ### Evaluation Script
 
+The evaluation script is available on our GitHub repo! [Check it out](https://github.com/room-b007/test-medicina)!
+
+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:
+
+```bash
+conda create -n medschool-test python=3.10
+conda activate medschool-test
+```
+
 > [!NOTE]
-> We will release the evaluation script soon. Stay tuned!
+> 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:
+
+```bash
+pip install --upgrade -r requirements.txt
+```
 
-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-entrance-exams`.
+#### Running the Evaluation Script
 
 To run the evaluation, you can use the following command:
 
 ```bash
+# 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"
 
-lm_eval \
+# 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 medschool_entrance_exams_it_mc,medschool_entrance_exams_it_cloze \
+  --tasks $TASKS \
   --batch_size auto \
   --log_samples \
-  --output_path outputs/ \
-  --include tasks/medschool-entrance-exams/
+  --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 [lm-evaluation-harness](https://github.com/eleutherai/lm-evaluation-harness) repository for more details on how to use the library.
+Please, refer to the examples in `examples/evaluation` or [lm-evaluation-harness](https://github.com/eleutherai/lm-evaluation-harness) repository for more details on how to use the library.
+
+> [!NOTE]
+> 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 <u> maximum possible score is 90 = 60 * 1.5 </u>, when all the answers are correct, while the <u> minum score is -24 = 60 * -0.4 </u>, 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](https://www.miur.gov.it/)), 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](https://domande-ap.mur.gov.it/domande).
+The dataset is collected from the official Italian website of the Ministry of Education, University and Research ([MIUR](https://www.miur.gov.it/)), 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](https://domande-ap.mur.gov.it/domande).
 
 ### Composition
 
@@ -260,10 +320,23 @@ where:
 ```
 
 
+### 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:
 
-## Contributing
+```bash
+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).
 
-Contributions to this dataset are welcome! If you have additional tasks or domains that you would like to include, please submit a pull request.
+
+## Citation
+> [!NOTE]
+> We are currently writing a report on the dataset and will provide the citation information soon.
 
 
 ## License