Video-Reason/VBVR-MultiStep

1---2license: cc-by-4.03task_categories:4  - image-to-video5  - text-to-video6language:7  - en8size_categories:9  - 100K<n<1M10pretty_name: VBVR-MultiStep11tags:12  - video-reasoning13  - multi-step14  - long-horizon15  - image-to-video16  - training17---18 19# VBVR-MultiStep20 21The **~360k-sample programmatic training corpus** for long-horizon multi-step image-to-video (I2V) reasoning. Companion to the frozen [VBVR-MultiStep-Bench](https://huggingface.co/datasets/Video-Reason/VBVR-MultiStep-Bench) (180-instance evaluation split).22 23Part of the **VBVR (Very Big Video Reasoning Suite)** project: <https://video-reason.com>. See [Wang et al., ICML 2026](https://icml.cc/virtual/2026/poster/65709) for the parent suite.24 25## At a glance26 27| Property | Value |28|---|---|29| Tasks | **36** parameterized tasks (`Multi-01` … `Multi-36`) |30| Reasoning families | Navigation, Planning, CSP, Execution, Geometry, Physics |31| Total samples | **~360,000** (≈10k per task) |32| Total size | **~164 GB** |33| Format | Tar.gz shards (nested per-sample folders) + Parquet metadata |34| Shards | 7,200 (≈50 samples per shard) |35| License | CC-BY-4.0 |36 37## Repository layout38 39```40.41├── README.md42├── croissant.json                       # Croissant + RAI metadata43├── data/44│   ├── metadata.parquet                 # global index of all 360k samples45│   └── metadata_shards/46│       └── Multi-XX_<name>.parquet      # per-task metadata (36 files)47├── questions/                           # WebDataset shards48│   └── Multi-XX_<name>_NNNNN-NNNNN.tar.gz49│       └── (50 samples per shard, 5 files per sample, see "Sample format" below)50└── sample/                              # ~5 GB representative subset for quick inspection51    ├── data/metadata_shards/...52    └── questions/                        # 6 shards × 36 tasks = 216 shards53```54 55The `sample/` subdirectory is a 5 GB pre-curated subset (the first 300 samples of every task) for reviewers and quick experimentation. To pull it:56 57```bash58huggingface-cli download Video-Reason/VBVR-MultiStep \59    --repo-type dataset \60    --include "sample/**" \61    --local-dir ./vbvr-multistep-sample62```63 64## Sample format (inside each `.tar.gz` shard)65 66Each shard expands to a nested folder tree, identical in shape to the evaluation split:67 68```69Multi-XX_<name>_data-generator/70└── Multi-XX_<name>_data-generator_task/71    └── Multi-XX_<name>_data-generator_<id>/72        ├── first_frame.png         # conditioning frame73        ├── prompt.txt              # natural-language task contract74        ├── final_frame.png         # target endpoint (held-out at inference)75        ├── ground_truth.mp4        # reference rollout76        └── question_metadata.json  # seed, version, tolerances, task fields77```78 79Each shard contains 50 such instance folders. The five-artifact contract is identical to the evaluation split.80 81To extract:82 83```bash84tar xzf Multi-01_maze_shortest_path_data-generator_00000-00049.tar.gz85```86 87## Loading88 89### Per-task metadata (recommended entry point)90 91```python92import pandas as pd93m = pd.read_parquet(94    "hf://datasets/Video-Reason/VBVR-MultiStep/data/metadata_shards/Multi-01_maze_shortest_path_data-generator.parquet"95)96print(m.head())97```98 99### Direct shard download100 101```python102from huggingface_hub import hf_hub_download103import tarfile104shard = hf_hub_download(105    "Video-Reason/VBVR-MultiStep",106    "questions/Multi-01_maze_shortest_path_data-generator_00000-00049.tar.gz",107    repo_type="dataset",108)109with tarfile.open(shard) as t:110    t.extractall("./extracted")111```112 113### Pull only the 5 GB sample114 115```bash116huggingface-cli download Video-Reason/VBVR-MultiStep \117    --repo-type dataset \118    --include "sample/**" \119    --local-dir ./vbvr-multistep-sample120```121 122## Splits and seeds123 124The training corpus is partitioned into disjoint seed bands:125 126| Band | Seed range | Samples per task | Total samples |127|---|---|---|---|128| First-half | 1–5,000 | 5,000 | ~170k (across 34 trained tasks) |129| Second-half | 5,001–10,000 | 5,000 | ~170k (across 34 trained tasks) |130 131Both bands are disjoint from the **180-instance evaluation seeds** in `VBVR-MultiStep-Bench`. The submitted paper trains on 34 of 36 tasks; the released corpus contains all 36 task families.132 133## Reasoning families134 135See the [bench dataset card](https://huggingface.co/datasets/Video-Reason/VBVR-MultiStep-Bench) for the family taxonomy. Each family contributes 6 tasks, for 36 total.136 137## Intended use and out-of-scope138 139- **Primary use**: training I2V systems on long-horizon multi-step reasoning under explicit per-step rules.140- **Out-of-scope**: this corpus is fully synthetic and stylized; transfer to unconstrained open-world video is not validated by this release.141- **Not validated for**: production VLM pretraining at scale, real-world video generation, or any safety-critical use.142 143## License144 145Released under **CC-BY-4.0**. Generators consume only released task definitions; no third-party copyrighted content is embedded.146 147Derivatives of `Wan2.2-I2V-A14B` (Apache-2.0) referenced in the companion paper comply with the upstream license. This dataset does not redistribute model weights.148 149## Responsible AI150 151The dataset is fully synthetic. There are no human subjects, no scraped media, and no personal or sensitive information. Known biases inherit from the deterministic generators — every task family covers a deliberately narrow conceptual slice, and visual style is controlled by a fixed renderer family (no demographic content). See [`croissant.json`](./croissant.json) for the complete RAI metadata.152 153