Convert dataset to new format with metadata.jsonl and updated image names

convert_format.py

@@ -0,0 +1,56 @@
+import json
+import os
+import shutil
+from pathlib import Path
+
+def convert_json_to_jsonl(input_file, output_file, split_name):
+    with open(input_file, 'r', encoding='utf-8') as f:
+        data = json.load(f)
+    
+    with open(output_file, 'w', encoding='utf-8') as f:
+        for item in data:
+            # Créer le nouveau format
+            new_item = {
+                "text": f"### Analyse du Document Financier\n\n#### 1. Question\n{item['query']}\n\n#### 2. Réponse\n{item['answer']}",
+                "metadata": {
+                    "document_id": f"KID_{item['image'].split('_')[-1].split('.')[0]}",
+                    "language": item['langue'],
+                    "source": "KID Document",
+                    "page_number": item['page'] if item['page'] else 1
+                },
+                "query": item['query'],
+                "file_name": f"images/output_page_KID_{item['image'].split('_')[-1].split('.')[0]}.jpg"
+            }
+            f.write(json.dumps(new_item, ensure_ascii=False) + '\n')
+
+def rename_images(split_dir):
+    images_dir = os.path.join(split_dir, 'images')
+    for img_file in os.listdir(images_dir):
+        if img_file.endswith('.png'):
+            new_name = f"output_page_KID_{img_file.split('_')[-1].split('.')[0]}.jpg"
+            old_path = os.path.join(images_dir, img_file)
+            new_path = os.path.join(images_dir, new_name)
+            shutil.copy2(old_path, new_path)
+            os.remove(old_path)
+
+def process_split(split_name):
+    split_dir = Path(split_name)
+    input_file = split_dir / f"{split_name}.json"
+    output_file = split_dir / "metadata.jsonl"
+    
+    # Convertir le JSON en JSONL
+    convert_json_to_jsonl(input_file, output_file, split_name)
+    
+    # Renommer les images
+    rename_images(split_dir)
+    
+    # Supprimer l'ancien fichier JSON
+    os.remove(input_file)
+
+def main():
+    # Traiter chaque split
+    for split in ['train', 'validation', 'test']:
+        process_split(split)
+
+if __name__ == "__main__":
+    main() 

dataset.yaml

@@ -1,8 +1,7 @@
 builder_type: imagefolder
 format: imagefolder
 license: mit
-language:
-  - fr
+language: - fr
 tags:
   - financial-documents
   - vision-language
@@ -10,18 +9,16 @@ tags:
 task_categories:
   - text-to-image
   - image-to-text
-splits:
-  train: train/train.json
-  validation: validation/validation.json
-  test: test/test.json
 features:
-  image:
-    type: Image
+  text:
+    dtype: string
+  metadata:
+    dtype: dict
   query:
-    type: string
-  answer:
-    type: string
-  langue:
-    type: string
-  page:
-    type: string 
+    dtype: string
+  file_name:
+    dtype: string
+splits:
+  train: train/metadata.jsonl
+  validation: validation/metadata.jsonl
+  test: test/metadata.jsonl