README.md

---
configs:
- config_name: size_variation
  data_files: size_variation.tsv
  sep: "\t"
  default: true
- config_name: phase_variation
  data_files: phase_variation.tsv
  sep: "\t"
- config_name: contrastive_probing
  data_files: contrastive_probing.tsv
  sep: "\t"
task_categories:
- image-text-to-text
---

# SpatialTunnel

SpatialTunnel is a Blender-rendered diagnostic dataset for studying how vision-language models (VLMs) represent spatial relations internally. It was introduced in the paper: **[Why Far Looks Up: Probing Spatial Representation in Vision-Language Models](https://huggingface.co/papers/2605.30161)**.

The dataset is designed to isolate spatial shortcut biases (like the perspective bias where "higher in image means farther away") by removing background and perspective correlations present in natural images.

### Resources

- **Project page:** [https://cheolhong0916.github.io/whyfarlooksup.github.io/](https://cheolhong0916.github.io/whyfarlooksup.github.io/)
- **Contrastive-probing code:** [https://github.com/cheolhong0916/contrastive-probing](https://github.com/cheolhong0916/contrastive-probing)
- **SpatialTunnel generation code:** [https://github.com/cube-c/spatialtunnel-dataset-gen](https://github.com/cube-c/spatialtunnel-dataset-gen)

## Dataset Configs

| Config | File | Rows | Description |
|---|---|---|---|
| `size_variation` | `size_variation.tsv` | 1,100 | Apparent-size questions under controlled object-size variation. |
| `phase_variation` | `phase_variation.tsv` | 3,072 | Distance questions with controlled angular-position variation. |
| `contrastive_probing` | `contrastive_probing.tsv` | 1,200 | Balanced spatial-relation questions for contrastive probing. |

## Format

All configs are tab-separated files with the following columns:

| Column | Description |
|---|---|
| `index` | Row index within the selected config. |
| `image` | Base64-encoded PNG image. |
| `question` | Spatial question to ask the model. |
| `answer` | Ground-truth answer for the row. |

## Sample Usage

You can download specific parts of the benchmark using the Hugging Face CLI:

```bash
huggingface-cli download cubec/spatialtunnel contrastive_probing.tsv --repo-type dataset --local-dir ./data
```

---

## Citation

If you use this dataset, please cite our paper:

```bibtex
@article{min2026whyfarlooksup,
  title   = {Why Far Looks Up: Probing Spatial Representation in Vision-Language Models},
  author  = {Min, Cheolhong and Jung, Jaeyun and Lee, Daeun and Jeon, Hyeonseong and
             Su, Yu and Tremblay, Jonathan and Song, Chan Hee and Park, Jaesik},
  journal = {arXiv preprint arXiv:2605.30161},
  year    = {2026},
}
```