Update README.md

README.md

@@ -88,6 +88,16 @@ results = model("path/to/road_image.jpg")
 results[0].show()
 ```
 
+## Training Data
+
+Trained on a road-damage detection dataset hosted on Roboflow (7 damage classes, exported in YOLOv8 format). Bring your own Roboflow API key and project reference to re-download the exact split used in `Road_Damage.ipynb`, or substitute any YOLO-format road-damage dataset with a matching `data.yaml`.
+
+## Training Procedure
+
+- **Image size:** 640×640 · **Batch size:** 16 · **Epochs:** up to 50 (early stopping)
+- **Early stopping patience:** 10 epochs for YOLOv8n, 3 epochs for YOLOv10n and YOLO11n
+- Each model trained independently from its official Ultralytics pretrained weights (`yolov8n.pt`, `yolov10n.pt`, `yolo11n.pt`)
+- Evaluated with the built-in Ultralytics validator (`model.val()`) — box precision/recall, mAP50, mAP50-95, and inference speed all reported directly from `DetMetrics`
 
 # Gradio User Interface
 
@@ -143,17 +153,6 @@ The Gradio interface can be deployed easily on:
 
 ---
 
-## Training Data
-
-Trained on a road-damage detection dataset hosted on Roboflow (7 damage classes, exported in YOLOv8 format). Bring your own Roboflow API key and project reference to re-download the exact split used in `Road_Damage.ipynb`, or substitute any YOLO-format road-damage dataset with a matching `data.yaml`.
-
-## Training Procedure
-
-- **Image size:** 640×640 · **Batch size:** 16 · **Epochs:** up to 50 (early stopping)
-- **Early stopping patience:** 10 epochs for YOLOv8n, 3 epochs for YOLOv10n and YOLO11n
-- Each model trained independently from its official Ultralytics pretrained weights (`yolov8n.pt`, `yolov10n.pt`, `yolo11n.pt`)
-- Evaluated with the built-in Ultralytics validator (`model.val()`) — box precision/recall, mAP50, mAP50-95, and inference speed all reported directly from `DetMetrics`
-
 ## Limitations
 
 This is a research/benchmarking project, not a production-ready inspection system. Detection quality (mAP50-95 in the 0.19–0.23 range) reflects a lightweight "nano" model family trained for a limited number of epochs on a single dataset — expect false negatives on damage types underrepresented in training data, and re-validate before any real-world deployment (e.g. road inspection, insurance assessment).