Update README.md

README.md

@@ -12,15 +12,15 @@ annotations_creators:
 size_categories:
   - 10K<n<100K
 tags:
+  - minecraft
   - computer-vision
   - multi-label-classification
   - image-feature-extraction
+  - gameplay
   - representation-learning
-  - embodied-ai
   - agent-perception
-  - minecraft
   - reinforcement-learning
-  - gameplay
+  - embodied-ai
 configs:
   - config_name: core
     default: true
@@ -172,14 +172,114 @@ dataset-root/
 └── splits/
 ```
 
-Load a local image:
+### Load images from the Hub
+
+The CSV configurations expose `image_path` as a relative string. Image files live under:
+
+```text
+images/<image_path>
+```
+
+#### Download one image
+
+Use `hf_hub_download()` for inspection, notebooks, or other cases where only a few images are needed:
+
+```python
+from datasets import load_dataset
+from huggingface_hub import hf_hub_download
+from PIL import Image
+
+REPO_ID = "OWNER/REPOSITORY"
+
+dataset = load_dataset(
+    REPO_ID,
+    "core",
+    split="train",
+)
+
+example = dataset[0]
+
+local_path = hf_hub_download(
+    repo_id=REPO_ID,
+    repo_type="dataset",
+    filename=f"images/{example['image_path']}",
+)
+
+image = Image.open(local_path).convert("RGB")
+
+print(example["image_path"])
+print(image.size)
+```
+
+Downloaded files are stored in the Hugging Face cache and are normally reused on later calls.
+
+#### Download the image store for training
+
+For training or large-scale feature extraction, download the image tree once with `snapshot_download()`:
+
+```python
+from pathlib import Path
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from PIL import Image
+
+REPO_ID = "OWNER/REPOSITORY"
+
+repo_root = Path(
+    snapshot_download(
+        repo_id=REPO_ID,
+        repo_type="dataset",
+        allow_patterns=["images/**"],
+    )
+)
+
+dataset = load_dataset(
+    REPO_ID,
+    "core",
+    split="train",
+)
+
+example = dataset[0]
+
+image = Image.open(
+    repo_root / "images" / example["image_path"]
+).convert("RGB")
+```
+
+The original directory structure is preserved inside the local snapshot.
+
+#### Add a decoded `image` column
+
+After downloading the image store, the string paths can be exposed as a lazy `datasets.Image` column:
+
+```python
+from datasets import Image
+
+def attach_image_path(example):
+    return {
+        "image": str(
+            repo_root
+            / "images"
+            / example["image_path"]
+        )
+    }
+
+dataset = dataset.map(attach_image_path)
+dataset = dataset.cast_column("image", Image())
+
+image = dataset[0]["image"]
+```
+
+`dataset[0]["image"]` then returns a decoded Pillow image, while `image_path` remains available as the stable dataset identifier.
+
+### Load images from a local checkout
 
 ```python
 from pathlib import Path
 from PIL import Image
 
 root = Path("/path/to/dataset-root")
-example = dataset["train"][0]
+example = dataset[0]
 
 image = Image.open(
     root / "images" / example["image_path"]