README.md

---
license: cc-by-4.0
task_categories:
  - image-to-video
  - text-to-video
language:
  - en
size_categories:
  - n<1K
pretty_name: VBVR-MultiStep-Bench
tags:
  - video-reasoning
  - multi-step
  - long-horizon
  - image-to-video
  - evaluation
  - benchmark
---

# VBVR-MultiStep-Bench

The frozen **180-instance public evaluation split** released alongside the [VBVR-MultiStep](https://huggingface.co/datasets/Video-Reason/VBVR-MultiStep) training corpus. Designed for long-horizon multi-step image-to-video (I2V) reasoning evaluation.

This dataset is part of the **VBVR (Very Big Video Reasoning Suite)** project. See the parent suite at <https://video-reason.com> and the suite paper [VBVR: A Very Big Video Reasoning Suite (Wang et al., ICML 2026)](https://icml.cc/virtual/2026/poster/65709).

## At a glance

| Property | Value |
|---|---|
| Tasks | **36** parameterized tasks (`Multi-01` … `Multi-36`) |
| Reasoning families | Navigation, Planning, CSP, Execution, Geometry, Physics |
| Instances | **180** (5 per task × 36) |
| Per-instance artifacts | 5 (see below) |
| License | CC-BY-4.0 |

## Five-artifact data contract

Every instance lives at:

```
Multi-XX_<name>_data-generator/Multi-XX_<name>_data-generator_task/Multi-XX_<name>_data-generator_<id>/
```

and contains exactly:

| File | Role |
|---|---|
| `first_frame.png` | Model conditioning image (the only visual input the model receives at inference) |
| `prompt.txt` | Natural-language task contract |
| `final_frame.png` | Target endpoint (held out from the model) |
| `ground_truth.mp4` | Reference rollout demonstrating the correct trajectory |
| `question_metadata.json` | Seed, version, tolerances, task-specific fields |

A top-level `metadata.parquet` indexes every instance with the task id, family, seed, and per-instance metadata for fast filtering.

## Reasoning families

| Family | Characteristic | Released tasks |
|---|---|---|
| Navigation | Discrete motion under adjacency / obstacle constraints | 6 |
| Planning | Operator-based state transformation | 6 |
| CSP | Incremental labeling under global consistency | 6 (3 used for human judging) |
| Execution | Clocked deterministic update rules | 6 |
| Geometry | Ordered constructive geometry | 6 |
| Physics | Continuous dynamics with contact / conservation | 6 |

Tasks `Multi-13`, `Multi-14`, `Multi-15` (CSP) are excluded from the human-judging pool described in the paper but are included in this release for completeness.

## Intended use

- **Primary use**: trajectory-level evaluation of I2V systems under a fixed five-artifact contract.
- **Comparison protocol**: blind human pairwise judging on three independent axes — process correctness, reference fidelity, render quality.
- **Companion training corpus**: [Video-Reason/VBVR-MultiStep](https://huggingface.co/datasets/Video-Reason/VBVR-MultiStep) (~360k samples).

## Loading

```python
import pandas as pd
meta = pd.read_parquet("hf://datasets/Video-Reason/VBVR-MultiStep-Bench/metadata.parquet")
```

Or pull a single instance:

```python
from huggingface_hub import hf_hub_download
prompt_path = hf_hub_download(
    "Video-Reason/VBVR-MultiStep-Bench",
    "Multi-01_maze_shortest_path_data-generator/Multi-01_maze_shortest_path_data-generator_task/Multi-01_maze_shortest_path_data-generator_00000000/prompt.txt",
    repo_type="dataset",
)
```

## License

Released under **CC-BY-4.0**. The reference rollouts are produced from generators that consume only released task definitions; no third-party copyrighted content is embedded.

Wan2.2-I2V-A14B (Apache-2.0) is referenced as a baseline model and a fine-tuning ancestor for `VBVR-Wan2.2`; this dataset does not redistribute Wan2.2 weights.

## Responsible AI

This dataset is fully synthetic — generators produce every instance from controlled parameters. There are no human subjects, no scraped media, and no personal information. See the [Croissant file](./croissant.json) for the complete RAI metadata.