SpatialTunnel is a Blender-rendered diagnostic dataset for studying how vision-language models represent spatial relations internally. It was introduced in Why Far Looks Up: Probing Spatial Representation in Vision-Language Models (arXiv:2605.30161).
| Config | File | Rows | Description |
|---|---|---|---|
phase_variation | phase_variation-*.parquet | 12,288 | Binary depth-comparison questions with controlled angular-position variation. |
size_variation | size_variation-*.parquet | 4,400 | Binary depth-comparison questions under controlled object-size variation. |
contrastive_probing | contrastive_probing.parquet | 1,200 | Balanced spatial-relation questions for contrastive probing. |
All configurations use Parquet files with the following columns:
For phase_variation and size_variation, each rendered image is paired with four binary VQA questions that vary the object order and comparison direction (closer/farther).
| Column | Description |
|---|---|
index | Row index within the selected config. |
image | PNG image stored as a Hugging Face Image feature. |
question | Spatial question to ask the model. |
answer | Ground-truth answer for the row. |
contrastive_probing.tsv is a TSV file for compatibility with the contrastive-probing code.
If you use this dataset, please cite our paper.
@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},
}
SpatialTunnel is a Blender-rendered diagnostic dataset for studying how vision-language models represent spatial relations internally. It was introduced in Why Far Looks Up: Probing Spatial Representation in Vision-Language Models (arXiv:2605.30161).
| Config | File | Rows | Description |
|---|---|---|---|
phase_variation | phase_variation-*.parquet | 12,288 | Binary depth-comparison questions with controlled angular-position variation. |
size_variation | size_variation-*.parquet | 4,400 | Binary depth-comparison questions under controlled object-size variation. |
contrastive_probing | contrastive_probing.parquet | 1,200 | Balanced spatial-relation questions for contrastive probing. |
All configurations use Parquet files with the following columns:
For phase_variation and size_variation, each rendered image is paired with four binary VQA questions that vary the object order and comparison direction (closer/farther).
| Column | Description |
|---|---|
index | Row index within the selected config. |
image | PNG image stored as a Hugging Face Image feature. |
question | Spatial question to ask the model. |
answer | Ground-truth answer for the row. |
contrastive_probing.tsv is a TSV file for compatibility with the contrastive-probing code.
If you use this dataset, please cite our paper.
@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},
}