README.md

---
license: mit
task_categories:
- visual-question-answering
- depth-estimation
language:
- en
tags:
- depth
- spatial-reasoning
- computer-vision
- multimodal
- vlm
- ablate-to-validate
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image_filename
    dtype: string
  - name: original_image
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  - name: answer_letter
    dtype: string
  - name: answer_type
    dtype: string
  - name: num_points
    dtype: int64
  - name: point_labels
    sequence: string
  - name: depth_token
    dtype: string
  - name: point_A_x
    dtype: int64
  - name: point_A_y
    dtype: int64
  - name: point_B_x
    dtype: int64
  - name: point_B_y
    dtype: int64
  - name: point_C_x
    dtype: int64
  - name: point_C_y
    dtype: int64
  - name: point_D_x
    dtype: int64
  - name: point_D_y
    dtype: int64
  - name: point_E_x
    dtype: int64
  - name: point_E_y
    dtype: int64
  - name: point_A_depth
    dtype: float64
  - name: point_B_depth
    dtype: float64
  - name: point_C_depth
    dtype: float64
  - name: point_D_depth
    dtype: float64
  - name: point_E_depth
    dtype: float64
  - name: image
    dtype: image
  splits:
  - name: train
    num_bytes: 466678289.0
    num_examples: 19279
  download_size: 421666211
  dataset_size: 466678289.0
pretty_name: 'Mixed-Depth: Relative-Depth Point QA (ADE20K)'
---

# Mixed-Depth: Relative-Depth Point QA (ADE20K)

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)).

Each example shows an ADE20K image with **3, 4, or 5 labeled points** circled and asks **which point is
closest to the camera**. 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 &middot; **Source imagery:** ADE20K (train split)

## Browse / load (parquet, default config)

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
ds = load_dataset("agianbig/mixed_depth", split="train")
ds[0]["image"]          # PIL.Image (336x336)
ds[0]["answer_letter"]  # e.g. "C"
```

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`.

> **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.

## 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
# 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` &mdash; long chain-of-thought answers (coords + depth reasoning + final letter)
- `mixed_depth_short.json` &mdash; short answers (final letter only, e.g. `(C)`)

## License

Released under MIT. Imagery derived from ADE20K; please also observe the ADE20K terms.