---
license: other
task_categories:
- image-to-text
language:
- en
tags:
- spatial-reasoning
- video-understanding
- world-model
---

# SpaVoBench

SpaVoBench is a timestamped image-and-instruction release derived from the human-reviewed `keep` records in WorldModelBench. Each JSONL row corresponds to exactly one image and one English caption.

## Layout

- `data/annotations.jsonl`: release records, one per extracted frame.
- `data/annotations.json`: the same keep-only records as a standard JSON array.
- `data/manifest.json`: release counts, source policy, and timestamp convention.
- `images/`: JPEG frames named by `annotation_id`.
- `source-videos/README.md` and `source-videos/manifest.jsonl`: canonical source-video policy and grouped source links. Complete MP4s remain in `Xinyun-Liu/WorldModelBench` and are not duplicated here.

## Timestamp Convention

One- and two-digit timestamps are seconds. Three- and four-digit timestamps use the final two digits as seconds and the leading digits as minutes. For example, `515` is 5:15 (315 seconds), and `0158` is 1:58 (118 seconds). `timestamp_raw` preserves the submitted value and `timestamp_seconds` contains the decoded extraction time.

## Provenance

Every record contains the original WorldModelBench file path and public URL. `text_caption` preserves the English instruction derived from the original annotation. `prompt_fields` stores the review-approved perspective, scene context, motion subject, camera constraint, and intervention separately; `model_prompt` is the fluent public caption composed only from perspective, camera constraint, and intervention. The release does not duplicate source MP4 files, so the source dataset's terms and provenance remain attached to the original video assets.