croissant.json

{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "cr": "http://mlcommons.org/croissant/",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "data": {"@id": "cr:data", "@type": "@json"},
    "dataType": {"@id": "cr:dataType", "@type": "@vocab"},
    "dct": "http://purl.org/dc/terms/",
    "examples": {"@id": "cr:examples", "@type": "@json"},
    "extract": "cr:extract",
    "field": "cr:field",
    "fileProperty": "cr:fileProperty",
    "fileObject": "cr:fileObject",
    "fileSet": "cr:fileSet",
    "format": "cr:format",
    "includes": "cr:includes",
    "isLiveDataset": "cr:isLiveDataset",
    "jsonPath": "cr:jsonPath",
    "key": "cr:key",
    "md5": "cr:md5",
    "parentField": "cr:parentField",
    "path": "cr:path",
    "recordSet": "cr:recordSet",
    "references": "cr:references",
    "regex": "cr:regex",
    "repeated": "cr:repeated",
    "replace": "cr:replace",
    "sc": "https://schema.org/",
    "separator": "cr:separator",
    "source": "cr:source",
    "subField": "cr:subField",
    "transform": "cr:transform"
  },
  "@type": "sc:Dataset",
  "name": "VBVR-MultiStep",
  "conformsTo": "http://mlcommons.org/croissant/1.0",
  "description": "The ~360,000-sample programmatic training corpus for long-horizon multi-step image-to-video reasoning. 36 parameterized tasks across six reasoning families (Navigation, Planning, CSP, Execution, Geometry, Physics). Distributed as 7,200 tar.gz shards (≈50 samples per shard) plus Parquet metadata; each instance follows a five-artifact contract identical to the VBVR-MultiStep-Bench evaluation split.",
  "alternateName": ["VBVR-MultiStep Training Corpus"],
  "creator": {
    "@type": "sc:Organization",
    "name": "Video-Reason",
    "url": "https://video-reason.com"
  },
  "datePublished": "2026-05-06",
  "keywords": ["video reasoning", "multi-step reasoning", "long-horizon", "image-to-video", "training", "synthetic", "tar.gz", "parquet"],
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "url": "https://huggingface.co/datasets/Video-Reason/VBVR-MultiStep",
  "version": "1.0.0",
  "isLiveDataset": false,
  "rai:dataCollection": "Fully synthetic. Each of the 36 tasks ships a deterministic generator that emits the five-artifact contract (first_frame.png, prompt.txt, final_frame.png, ground_truth.mp4, question_metadata.json) from a (task, seed) pair. No scraping, no human subjects, no third-party media, no manual annotation. The training corpus is partitioned into disjoint seed bands (1–5,000 and 5,001–10,000 per task) that are themselves disjoint from the evaluation seeds released in VBVR-MultiStep-Bench.",
  "rai:dataCollectionType": ["Synthetic"],
  "rai:hasSyntheticData": true,
  "rai:dataPreprocessingProtocol": "Per-task generator output is grouped into 50-sample batches and packed into tar.gz shards under questions/. Per-task Parquet metadata files (data/metadata_shards/) and a global metadata.parquet index every instance with task id, family, seed, and per-task fields. No sample is filtered, dropped, or transformed after generation.",
  "rai:dataAnnotationProtocol": "No human annotation. ground_truth.mp4 is rendered by each task's deterministic ground-truth solver.",
  "rai:dataAnnotationPlatform": "N/A (no annotation).",
  "rai:dataReleaseMaintenancePlan": "Versioned releases on the Hugging Face Hub. The Croissant file in this repository is the canonical long-term record. A 5 GB representative subset is provided under sample/ for quick inspection and reviewer convenience.",
  "rai:dataLimitations": [
    "Synthetic and stylized: transfer to unconstrained open-world video is not validated.",
    "Visual rendering is intentionally simplified to keep the symbolic state recoverable from frames; this is not a photorealism corpus.",
    "Per-task generator parameter ranges are bounded (e.g., maze sizes, planning horizons, physics regimes); the corpus does not span the long tail of any single family.",
    "Reference rollouts encode one valid trajectory per instance; alternative valid trajectories are not enumerated.",
    "Although 36 tasks ship, only 34 are used in the training experiments described in the companion paper; this release contains all 36 task families."
  ],
  "rai:dataBiases": [
    "Family balance is uniform (6 tasks per family) by design and does not reflect natural prevalence of these reasoning patterns.",
    "Generator parameters bias the difficulty distribution toward bounded and seed-controlled regimes that are amenable to symbolic ground truth; rare or open-ended cases are out of scope.",
    "Visual style is monocular, planar, and rendered by a fixed family of renderers; appearance distribution does not approximate any real-world video corpus and does not contain demographic content.",
    "No human demographic information is generated; bias along human demographic axes does not apply."
  ],
  "rai:personalSensitiveInformation": "None. The dataset contains no personal information, no biometric data, no demographic information, and no human subjects. All visual content is procedurally generated geometric, symbolic, or physical scenes.",
  "rai:dataUseCases": "Training image-to-video systems on long-horizon multi-step reasoning under explicit per-step rules. Validated use case in the companion paper: fine-tuning Wan2.2-I2V-A14B (Apache-2.0) with Dual-DiT two-phase LoRA. Out-of-scope: production VLM pretraining at scale, real-world video generation, or any safety-critical use.",
  "rai:dataSocialImpact": "Intended for academic research on reasoning evaluation in video generation. Risks are minimal: the dataset is synthetic, free of personal content, and rendered in a stylized regime not representative of any real population. The most plausible concern is research-direction effects (e.g., over-investing in stylized synthetic benchmarks), which we mitigate by positioning this corpus as a complement to (not a replacement for) appearance-centric and real-world video corpora.",
  "rai:dataReleaseUpdate": "If post-release errors are discovered, fixes will be published as additive shards or replacement Parquet entries in a new dataset version, with the prior version retained at its commit hash for backwards reproducibility."
}