diff --git a/main.py b/main.py
index 66516b6..f1c2162 100644
--- a/main.py
+++ b/main.py
@@ -7,7 +7,6 @@ import requests
 import os
 from multiprocessing import Pool
 
-
 CACHE_FOLDER = "cache/fonts"
 
 # Créer le dossier de cache s'il n'existe pas
@@ -104,19 +103,21 @@ def create_letter_image(letter, output_path, font_path):
 def createImageForEachLetter(args):
     """
     Function to create an image for each of the letters passed as an argument
-    :param args: Tuple containing output folder, font path, and index
+    :param args: Tuple containing output folder, font path, index, and type (training/validation)
     """
-    output_folder, font_path, index = args
+    output_folder, font_path, index, dataset_type = args
     alphabet = string.ascii_letters  # Inclut à la fois les lettres minuscules et majuscules
 
-    os.makedirs(output_folder, exist_ok=True)  # Ensure base output folder exists
+    base_folder = os.path.join(output_folder, dataset_type)
+    os.makedirs(base_folder, exist_ok=True)
 
     for element in alphabet:
-        letter_folder = os.path.join(output_folder, element)
+        letter_folder = os.path.join(base_folder, element)
         os.makedirs(letter_folder, exist_ok=True)
 
         create_letter_image(element, f"{letter_folder}/{element}-{index}.png", font_path)
 
+
 if __name__ == "__main__":
     if len(argv) < 3:
         raise Exception("Usage: " + argv[0] + " <number-of-iterations> <api-key>")
@@ -132,9 +133,13 @@ if __name__ == "__main__":
     # Fetch fonts and cache them
     fonts = [getRandomFont(api_key, usedFont) for _ in range(num_iterations)]
 
-    # Prepare arguments for multiprocessing
-    args = [("dataset", fonts[i], i) for i in range(num_iterations)]
+    # Split into training and validation datasets
+    num_training = int(num_iterations * 0.7)
+    num_validation = num_iterations - num_training
+
+    training_args = [("dataset", fonts[i], i, "training") for i in range(num_training)]
+    validation_args = [("dataset", fonts[i], i, "validation") for i in range(num_training, num_iterations)]
 
     # Use multiprocessing pool
     with Pool() as pool:
-        pool.map(createImageForEachLetter, args)
+        pool.map(createImageForEachLetter, training_args + validation_args)
\ No newline at end of file