Write dataset card

README.md

@@ -10,8 +10,10 @@ tags:
 - spatial-reasoning
 - computer-vision
 - multimodal
+- vlm
+- ablate-to-validate
 size_categories:
-- 1K<n<10K
+- 10K<n<100K
 configs:
 - config_name: default
   data_files:
@@ -77,12 +79,52 @@ dataset_info:
     num_examples: 19279
   download_size: 421666211
   dataset_size: 466678289.0
+pretty_name: 'Mixed-Depth: Relative-Depth Point QA (ADE20K)'
 ---
 
-# Depth Point Dataset
+# Mixed-Depth: Relative-Depth Point QA (ADE20K)
 
-This dataset contains 19279 depth-based question-answer pairs for training vision-language models.
+Training data for **Ablate-to-Validate: Are Vision-Language Models Really Using Visual Reasoning Tokens?**
+([code](https://github.com/tjazhang/ablate_to_validate) &middot; [project page](https://tjazhang.github.io/ablate_to_validate/) &middot; [arXiv](https://arxiv.org/abs/2605.21642)).
 
-- **Total samples**: 19279
-- **Image resolution**: 336x336
-- **Task**: Identify which labeled point is closest to the camera based on depth
+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.
+
+- **Examples:** 19,279 &mdash; 3-point: 6,736 &middot; 4-point: 6,562 &middot; 5-point: 5,981
+- **Image resolution:** 336 &times; 336
+- **Source imagery:** ADE20K (train split)
+
+## Browse / load (parquet, default config)
+
+The default config powers the dataset viewer and loads directly:
+
+```python
+from datasets import load_dataset
+ds = load_dataset("agianbig/mixed_depth", split="train")
+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).
+
+## Finetuning files (LLaVA / Qwen conversation format)
+
+For training, use the conversation-format JSON plus the image folder. Image fields are **basenames**
+relative to `images/`:
+
+```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
+#   --image_folder mixed_depth/images
+```
+
+- `mixed_depth_long.json` &mdash; long chain-of-thought answers (point coordinates + depth reasoning + final letter)
+- `mixed_depth_short.json` &mdash; short answers (final letter only, e.g. `(C)`)
+- `images/` &mdash; 19,279 JPGs, referenced by basename in the JSON
+
+## License
+
+Released under MIT. Imagery is derived from ADE20K; please also observe the ADE20K terms.