Spatialtunnel

SpatialTunnel is a Blender-rendered diagnostic dataset introduced in the paper 'Why Far Looks Up: Probing Spatial Representation in Vision-Language Models'. It was created by author 'cubec' and last updated on the Hugging Face platform on 2026-05-29. The dataset is designed to study how vision-language models represent spatial relations internally.

Use Cases

• Probing internal model representations of spatial relations based on the diagnostic nature of the dataset.
• Benchmarking vision-language model performance on spatial reasoning tasks based on the described configs like 'size_variation'.
• Studying the effect of synthetic, controlled visual stimuli on model understanding based on the Blender-rendered generation method.

Strengths

• Dataset is specifically designed for diagnostic research on a defined problem: spatial representation in vision-language models.
• One known configuration ('size_variation') contains 1,100 rows, providing a concrete scale for that subset.
• Associated resources like generation code and probing code are provided, suggesting reproducibility.

Limitations

• Description metadata is limited; actual data quality, column definitions, and full dataset scale require manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The overall dataset size, file formats, and license information are unknown, which may limit suitability assessment.

Provenance

Source

Hugging Face dataset uploaded by author 'cubec'.

Collection Method

Synthetically generated using Blender rendering software, as stated in the description.

Freshness

Last updated 2026-05-29 06:35:33; freshness should be verified.