add example OCRBench dataset

examples/custom/OCRBench.py

@@ -0,0 +1,45 @@
+
+from typing import Any, Dict, Optional
+from swift.llm import (load_dataset, register_dataset, DatasetMeta)
+from swift.llm.dataset import ResponsePreprocessor
+
+
+
+class OCRBenchPreprocessor(ResponsePreprocessor):
+
+    def preprocess(self, row: Dict[str, Any]) -> Optional[Dict[str, Any]]:
+        import os
+        row = super().preprocess(row)
+        image = row['image_path']
+        if not image:
+            return
+        image = os.path.join(self.prefix_path, image)
+        if not os.path.exists(image):
+            return
+        row['images'] = [image]
+        return row
+
+    def prepare_dataset(self, dataset):
+        "local dataset"
+
+        self.prefix_path = "/datasets/OCRBench/OCRBench_Images/"
+        return super().prepare_dataset(dataset)
+
+
+register_dataset(
+    DatasetMeta(
+        dataset_path='/datasets/OCRBench/OCRBench.json',
+        preprocess_func=OCRBenchPreprocessor(),
+
+    ))
+
+
+
+
+if __name__ == '__main__':
+    # The Shell script can view `examples/pytorch/llm/scripts/custom`.
+    # test dataset
+    train_dataset, val_dataset = load_dataset('/datasets/OCRBench/OCRBench.json')
+    print(f'train_dataset: {train_dataset}')
+    print(f'val_dataset: {val_dataset}')
+