Update README.md

README.md

@@ -19,10 +19,10 @@ All samples are stored in a Hugging Face-compatible parquet file at `data/<split
    ```python
    from datasets import load_dataset
 
-   ds = load_dataset('path/to/this/folder', split='train')
+   ds = load_dataset('path/to/this/folder', split='test')
    sample = ds[0]
    ``
-   `sample["preview"]` loads directly as a compressed PIL image; `sample["raw_image_path"]` points to the higher-quality file for inference.
+   `sample["raw_image_path"]` points to the higher-quality file for inference. `sample["preview"]` loads directly as a compressed PIL image.
 3. Use the metadata to drive evaluation logic, e.g., compute city-level accuracy, filter by `data_type`, or inspect specific regions.
 
 ## Notes