---
license: mit
task_categories:
- visual-question-answering
language:
- en
tags:
- spatial-reasoning
- 3d
- counting
- video
- benchmark
pretty_name: Spatial Scene Bench
size_categories:
- 1K<n<10K
dataset_info:
- config_name: full
  features:
  - name: id
    dtype: string
  - name: scene_id
    dtype: string
  - name: question_type
    dtype: string
  - name: modality
    dtype: string
  - name: question
    dtype: string
  - name: options
    list: string
  - name: answer
    dtype: string
  - name: answer_type
    dtype: string
  - name: images
    list: image
  splits:
  - name: test
    num_bytes: 290248515
    num_examples: 4387
  download_size: 259451429
  dataset_size: 290248515
- config_name: mini
  features:
  - name: id
    dtype: string
  - name: scene_id
    dtype: string
  - name: question_type
    dtype: string
  - name: modality
    dtype: string
  - name: question
    dtype: string
  - name: options
    list: string
  - name: answer
    dtype: string
  - name: answer_type
    dtype: string
  - name: images
    list: image
  splits:
  - name: test
    num_bytes: 17657343
    num_examples: 275
  download_size: 10124080
  dataset_size: 17657343
configs:
- config_name: full
  data_files:
  - split: test
    path: full/test-*
- config_name: mini
  data_files:
  - split: test
    path: mini/test-*
---

# Spatial Scene Bench

[Code](https://github.com/selfishout/spatial-scene-bench) | [lmms-eval task](https://github.com/EvolvingLMMs-Lab/lmms-eval)

Counting and relational reasoning over procedurally rendered 3D scenes, from single images to 16-frame orbit
videos. Every answer is computed from the 3D scene itself, and ambiguous questions (heavily occluded objects,
near-ties in depth or distance) are filtered out at generation time.

## Configs

| Config | Questions | Description |
| --- | --- | --- |
| `full` | 4,387 | 200 scenes, all 11 question types |
| `mini` | 275 | first 25 questions of each type |

```python
from datasets import load_dataset
ds = load_dataset("alithema/spatial-scene-bench", "mini", split="test")
```

## Fields

- `question_type`: one of `count`, `count_relation`, `relation`, `compare_count`, `closest`, `distance`,
  `ego_direction`, `video_count`, `video_closest`, `video_distance`, `video_ego_direction`
- `modality`: `image` (one 320×240 view) or `video` (16 orbit frames)
- `answer_type`: `count` (integer), `yesno`, `choice` (A–D, options in `options`), or `numeric` (meters)
- `images`: the view or the video frames, in order

## Scoring

Exact match for counts, yes/no and multiple choice; VSI-Bench Mean Relative Accuracy (MRA) for distances. The
headline score is the macro average over question types. Reference implementations: `spatial_scene_bench.evaluate`
in the code repository and the `spatial_scene_bench` task in lmms-eval.

| Model | Overall (mini) |
| --- | --- |
| Random baseline | 23.9 |
| Qwen2.5-VL-3B-Instruct (zero-shot) | 50.3 |

## Regenerating

The dataset is fully reproducible from the generator (seed 2025):

```bash
pip install "spatial-scene-bench[export] @ git+https://github.com/selfishout/spatial-scene-bench.git"
ssb-generate --out data/ssb --num-scenes 200 --video-frames 16 --seed 2025
python scripts/export_hf.py --data data/ssb --out data/ssb_hf
```