Mark7121983123/VR-MultiStep

1---2license: cc-by-4.03task_categories:4  - image-to-video5  - text-to-video6language:7  - en8size_categories:9  - 100K<n<1M10pretty_name: VR-MultiStep11tags:12  - video-reasoning13  - multi-step14  - long-horizon15  - image-to-video16  - training17---18 19# VR-MultiStep20 21The **~360k-sample programmatic training corpus** for long-horizon multi-step image-to-video (I2V) reasoning. Companion to the frozen **VR-MultiStep-Bench** (180-instance evaluation split).22 23## At a glance24 25| Property | Value |26|---|---|27| Tasks | **36** parameterized tasks (`Multi-01` … `Multi-36`) |28| Reasoning families | Navigation, Planning, CSP, Execution, Geometry, Physics |29| Total samples | **~360,000** (≈10k per task) |30| Total size | **~164 GB** |31| Format | Tar.gz shards (nested per-sample folders) + Parquet metadata |32| Shards | 7,200 (≈50 samples per shard) |33| License | CC-BY-4.0 |34 35## Repository layout36 37```38.39├── README.md40├── croissant.json                       # Croissant + RAI metadata41├── data/42│   ├── metadata.parquet                 # global index of all 360k samples43│   └── metadata_shards/44│       └── Multi-XX_<name>.parquet      # per-task metadata (36 files)45├── questions/                           # WebDataset shards46│   └── Multi-XX_<name>_NNNNN-NNNNN.tar.gz47│       └── (50 samples per shard, 5 files per sample, see "Sample format" below)48└── sample/                              # ~5 GB representative subset for quick inspection49    ├── data/metadata_shards/...50    └── questions/                        # 6 shards × 36 tasks = 216 shards51```52 53The `sample/` subdirectory is a 5 GB pre-curated subset (the first 300 samples of every task) for reviewers and quick experimentation. To pull it:54 55```bash56huggingface-cli download Mark7121983123/VR-MultiStep \57    --repo-type dataset \58    --include "sample/**" \59    --local-dir ./vr-multistep-sample60```61 62## Sample format (inside each `.tar.gz` shard)63 64Each shard expands to a nested folder tree, identical in shape to the evaluation split:65 66```67Multi-XX_<name>_data-generator/68└── Multi-XX_<name>_data-generator_task/69    └── Multi-XX_<name>_data-generator_<id>/70        ├── first_frame.png         # conditioning frame71        ├── prompt.txt              # natural-language task contract72        ├── final_frame.png         # target endpoint (held-out at inference)73        ├── ground_truth.mp4        # reference rollout74        └── question_metadata.json  # seed, version, tolerances, task fields75```76 77Each shard contains 50 such instance folders. The five-artifact contract is identical to the evaluation split.78 79To extract:80 81```bash82tar xzf Multi-01_maze_shortest_path_data-generator_00000-00049.tar.gz83```84 85## Loading86 87### Per-task metadata (recommended entry point)88 89```python90import pandas as pd91m = pd.read_parquet(92    "hf://datasets/Mark7121983123/VR-MultiStep/data/metadata_shards/Multi-01_maze_shortest_path_data-generator.parquet"93)94print(m.head())95```96 97### Direct shard download98 99```python100from huggingface_hub import hf_hub_download101import tarfile102shard = hf_hub_download(103    "Mark7121983123/VR-MultiStep",104    "questions/Multi-01_maze_shortest_path_data-generator_00000-00049.tar.gz",105    repo_type="dataset",106)107with tarfile.open(shard) as t:108    t.extractall("./extracted")109```110 111### Pull only the 5 GB sample112 113```bash114huggingface-cli download Mark7121983123/VR-MultiStep \115    --repo-type dataset \116    --include "sample/**" \117    --local-dir ./vr-multistep-sample118```119 120## Splits and seeds121 122The training corpus is partitioned into disjoint seed bands:123 124| Band | Seed range | Samples per task | Total samples |125|---|---|---|---|126| First-half | 1–5,000 | 5,000 | ~170k (across 34 trained tasks) |127| Second-half | 5,001–10,000 | 5,000 | ~170k (across 34 trained tasks) |128 129Both bands are disjoint from the **180-instance evaluation seeds** in **VR-MultiStep-Bench**. The submitted paper trains on 34 of 36 tasks; the released corpus contains all 36 task families.130 131## Reasoning families132 133See the [evaluation split dataset card](https://huggingface.co/datasets/Mark7121983123/VR-MultiStep-Bench) for the family taxonomy. Each family contributes 6 tasks, for 36 total.134 135## Intended use and out-of-scope136 137- **Primary use**: training I2V systems on long-horizon multi-step reasoning under explicit per-step rules.138- **Out-of-scope**: this corpus is fully synthetic and stylized; transfer to unconstrained open-world video is not validated by this release.139- **Not validated for**: production VLM pretraining at scale, real-world video generation, or any safety-critical use.140 141## License142 143Released under **CC-BY-4.0**. Generators consume only released task definitions; no third-party copyrighted content is embedded.144 145Derivatives 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.146 147## Responsible AI148 149The 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.150