cubec/spatialtunnel

1---2configs:3- config_name: phase_variation4  data_files: phase_variation-*.parquet5  default: true6- config_name: size_variation7  data_files: size_variation-*.parquet8- config_name: contrastive_probing9  data_files: contrastive_probing.parquet10dataset_info:11  features:12  - name: index13    dtype: int6414  - name: image15    dtype: image16  - name: question17    dtype: string18  - name: answer19    dtype: string20---21 22# SpatialTunnel23 24SpatialTunnel 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](https://arxiv.org/abs/2605.30161)).25 26## Resources27 28- [Project page](https://cheolhong0916.github.io/whyfarlooksup.github.io/)29- [Contrastive-probing code](https://github.com/cheolhong0916/contrastive-probing)30- [SpatialTunnel generation code](https://github.com/cube-c/spatialtunnel-dataset-gen)31 32## Dataset Configurations33 34| Config | File | Rows | Description |35|---|---|---|---|36| `phase_variation` | `phase_variation-*.parquet` | 12,288 | Binary depth-comparison questions with controlled angular-position variation. |37| `size_variation` | `size_variation-*.parquet` | 4,400 | Binary depth-comparison questions under controlled object-size variation. |38| `contrastive_probing` | `contrastive_probing.parquet` | 1,200 | Balanced spatial-relation questions for contrastive probing. |39 40## Format41 42All configurations use Parquet files with the following columns:43 44For `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`).45 46| Column | Description |47|---|---|48| `index` | Row index within the selected config. |49| `image` | PNG image stored as a Hugging Face `Image` feature. |50| `question` | Spatial question to ask the model. |51| `answer` | Ground-truth answer for the row. |52 53`contrastive_probing.tsv` is a TSV file for compatibility with the [contrastive-probing code](https://github.com/cheolhong0916/contrastive-probing).54 55---56 57## Citation58 59If you use this dataset, please cite our paper.60 61```bibtex62@article{min2026whyfarlooksup,63  title   = {Why Far Looks Up: Probing Spatial Representation in Vision-Language Models},64  author  = {Min, Cheolhong and Jung, Jaeyun and Lee, Daeun and Jeon, Hyeonseong and65             Su, Yu and Tremblay, Jonathan and Song, Chan Hee and Park, Jaesik},66  journal = {arXiv preprint arXiv:2605.30161},67  year    = {2026},68}69```70