Card: parquet-based images + extract-for-training snippet

README.md

@@ -88,15 +88,14 @@ Training data for **Ablate-to-Validate: Are Vision-Language Models Really Using
 ([code](https://github.com/tjazhang/ablate_to_validate) · [project page](https://tjazhang.github.io/ablate_to_validate/) · [arXiv](https://arxiv.org/abs/2605.21642)).
 
 Each example shows an ADE20K image with **3, 4, or 5 labeled points** circled and asks **which point is
-closest to the camera**. It is used to train the LLaVA and Qwen2.5-VL relative-depth models in the paper.
+closest to the camera**. Used to train the LLaVA and Qwen2.5-VL relative-depth models in the paper.
 
 - **Examples:** 19,279 — 3-point: 6,736 · 4-point: 6,562 · 5-point: 5,981
-- **Image resolution:** 336 × 336
-- **Source imagery:** ADE20K (train split)
+- **Image resolution:** 336 × 336 · **Source imagery:** ADE20K (train split)
 
 ## Browse / load (parquet, default config)
 
-The default config powers the dataset viewer and loads directly:
+All 19,279 images live (one per row, losslessly) in the parquet under `data/`, which also powers the viewer:
 
 ```python
 from datasets import load_dataset
@@ -105,26 +104,36 @@ ds[0]["image"]          # PIL.Image (336x336)
 ds[0]["answer_letter"]  # e.g. "C"
 ```
 
-Columns: `id`, `image`, `question`, `answer`, `answer_letter`, `answer_type`, `num_points`,
-`point_labels`, per-point `point_{A..E}_x` / `point_{A..E}_y` (pixel coords) and
-`point_{A..E}_depth` (relative depth).
+Columns: `id`, `image`, `image_filename`, `question`, `answer`, `answer_letter`, `answer_type`,
+`num_points`, `point_labels`, per-point `point_{A..E}_x` / `_y` (pixel coords) and `point_{A..E}_depth`.
 
-## Finetuning files (LLaVA / Qwen conversation format)
+> **Why no `images/` folder?** The Hugging Face Hub caps any directory at 10,000 files, and there are
+> 19,279 images. They are therefore distributed inside the parquet rather than as a flat folder.
 
-For training, use the conversation-format JSON plus the image folder. Image fields are **basenames**
-relative to `images/`:
+## Finetuning (LLaVA / Qwen conversation format)
+
+The conversation-format JSONs reference images by **basename**, so first extract the images to a local
+`images/` folder from the parquet, then train:
+
+```python
+import os
+from datasets import load_dataset
+ds = load_dataset("agianbig/mixed_depth", split="train")
+os.makedirs("images", exist_ok=True)
+for r in ds:
+    r["image"].save(os.path.join("images", os.path.basename(r["image_filename"])))
+```
 
 ```bash
 hf download agianbig/mixed_depth --repo-type dataset --local-dir mixed_depth
-# then, e.g.:
-#   --data_path   mixed_depth/mixed_depth_long.json   # or mixed_depth_short.json
+# run the snippet above (cwd = mixed_depth) to create mixed_depth/images/, then:
+#   --data_path    mixed_depth/mixed_depth_long.json   # CoT answers, or mixed_depth_short.json
 #   --image_folder mixed_depth/images
 ```
 
-- `mixed_depth_long.json` — long chain-of-thought answers (point coordinates + depth reasoning + final letter)
+- `mixed_depth_long.json` — long chain-of-thought answers (coords + depth reasoning + final letter)
 - `mixed_depth_short.json` — short answers (final letter only, e.g. `(C)`)
-- `images/` — 19,279 JPGs, referenced by basename in the JSON
 
 ## License
 
-Released under MIT. Imagery is derived from ADE20K; please also observe the ADE20K terms.
+Released under MIT. Imagery derived from ADE20K; please also observe the ADE20K terms.