Add dataset card (raw arrow, load_from_disk)

README.md

@@ -0,0 +1,70 @@
+---
+license: other
+task_categories:
+- visual-question-answering
+- image-to-text
+language:
+- en
+tags:
+- vision-flan
+- multimodal
+- instruction-tuning
+- vision-language
+size_categories:
+- 100K<n<1M
+---
+
+# Vision-FLAN (191-task) — Arrow (images embedded)
+
+The [Vision-FLAN](https://vision-flan.github.io/) `vision-flan_191-task_1k` visual instruction-tuning
+set converted to a 🤗 `datasets` **Arrow** dataset with **image bytes embedded**. 186,103 examples
+spanning ~191 human-labeled vision tasks.
+
+> ⚠️ This repo is a raw `save_to_disk` Arrow snapshot. The **dataset viewer** and
+> **`load_dataset()` do not work** here — load it with **`load_from_disk`** as shown below.
+
+## Loading
+
+```python
+from huggingface_hub import snapshot_download
+from datasets import load_from_disk
+
+path = snapshot_download("Ethlake/vision-flan", repo_type="dataset")
+ds = load_from_disk(path)
+print(ds)        # 186103 rows
+print(ds[0])
+```
+
+## Schema
+
+| column          | type                    | notes                                                       |
+|-----------------|-------------------------|-------------------------------------------------------------|
+| `images`        | list[Image]             | embedded bytes, decode to `PIL.Image` (typically 1 per row) |
+| `texts`         | list[{user, assistant}] | paired conversation turns (mostly single-turn)              |
+| `source_subset` | string                  | source task, e.g. `VQA_counting`, `HICO+human_activity_detection` |
+
+Example row:
+
+```python
+{
+ 'images': [<PIL.Image RGB 427x640>],
+ 'texts': [
+   {'user': 'This task tests your ability to count number of objects. '
+            'Here is the question "How many chairs are in the photo?".',
+    'assistant': 'two'},
+ ],
+ 'source_subset': 'VQA_counting',
+}
+```
+
+## Notes
+
+- ~191 distinct `source_subset` task types; filter on this column to slice by task.
+- Image bytes are embedded directly in the Arrow shards — self-contained, no separate image folders.
+- `images` is a list (multi-image capable) even though most examples contain a single image.
+
+## Source
+
+Derived from [Vision-Flan/vision-flan_191-task_1k](https://huggingface.co/datasets/Vision-Flan/vision-flan_191-task_1k)
+("Vision-Flan: Scaling Human-Labeled Tasks in Visual Instruction Tuning"). Please respect the
+original dataset's license and the licenses of its upstream image sources.