Add dataset card

README.md

@@ -1,3 +1,60 @@
 ---
-license: mit
+license: cc0-1.0
+task_categories:
+  - object-detection
+tags:
+  - welding
+  - defect-detection
+  - manufacturing
+  - coco
+  - yolo
+pretty_name: Welding Defect Object Detection (YOLO + COCO)
 ---
+
+# Welding Defect Object Detection
+
+2,028 annotated images of welds for defect detection, in **both YOLO and COCO
+formats**. Three classes:
+
+| id (YOLO / COCO) | name |
+|---|---|
+| 0 / 1 | Bad Weld |
+| 1 / 2 | Good Weld |
+| 2 / 3 | Defect |
+
+## Splits
+
+| split | images | annotations |
+|---|---|---|
+| train | 1,619 | 4,583 |
+| valid | 283 | 802 |
+| test | 126 | 301 |
+
+## Layout
+
+```
+├── data.yaml            # YOLO class names + split paths
+├── train|valid|test/
+│   ├── images/          # .jpg
+│   └── labels/          # YOLO .txt (class cx cy w h, normalized)
+└── coco/
+    ├── train.json       # COCO detection format
+    ├── valid.json
+    └── test.json
+```
+
+## COCO conversion notes
+
+The `coco/` jsons were generated from the YOLO labels with the
+[flux](https://github.com/rikkarth) YOLO→COCO converter:
+
+- bbox = `[x_min, y_min, width, height]`, float pixels
+- boxes clamped to image bounds
+- category ids are one-based (YOLO class 0 → COCO id 1)
+- annotation count parity verified: 5,686 YOLO label lines → 5,686 COCO annotations
+
+## Source & license
+
+Original dataset published on Kaggle by sukmaadhiwijaya as
+[Welding Defect - Object Detection](https://www.kaggle.com/datasets/sukmaadhiwijaya/welding-defect-object-detection)
+under **CC0: Public Domain**. This mirror adds the COCO-format annotations.