| { |
| "@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", |
| "prov": "http://www.w3.org/ns/prov#" |
| }, |
| "@type": "sc:Dataset", |
| "name": "SYNCR", |
| "description": "SYNCR is a synthetic cross-video reasoning benchmark for evaluating multimodal large language models on questions that require reasoning across multiple independent videos. The dataset contains 8,163 multiple-choice question-answer examples spanning four reasoning categories and eight tasks, with programmatically grounded annotations and video references.", |
| "conformsTo": "http://mlcommons.org/croissant/1.0", |
| "url": "https://huggingface.co/datasets/CrossVideoReasoning/SYNCR", |
| "license": "https://spdx.org/licenses/MIT.html", |
| "version": "1.0.0", |
| "datePublished": "2024-01-01", |
| "citeAs": "@article{syncr2026,\n title={SYNCR: A Cross-Video Reasoning Benchmark with Synthetic Grounding},\n author={Anonymous},\n year={2026}\n}", |
| "distribution": [ |
| { |
| "@type": "cr:FileSet", |
| "@id": "metadata_jsonl", |
| "name": "metadata_jsonl", |
| "description": "JSON Lines annotation file for the SYNCR test split.", |
| "encodingFormat": "application/jsonlines", |
| "includes": "test/metadata.jsonl" |
| } |
| ], |
| "recordSet": [ |
| { |
| "@type": "cr:RecordSet", |
| "@id": "default", |
| "name": "default", |
| "description": "SYNCR test split containing cross-video reasoning questions, multiple-choice options, answers, and video references.", |
| "field": [ |
| { |
| "@type": "cr:Field", |
| "@id": "default/category", |
| "name": "category", |
| "description": "High-level reasoning category, e.g., Temporal Alignment, Spatial Tracking, Comparative Reasoning, or Holistic Synthesis.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "metadata_jsonl" |
| }, |
| "extract": { |
| "jsonPath": "$.category" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/task", |
| "name": "task", |
| "description": "Specific benchmark task within the reasoning category.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "metadata_jsonl" |
| }, |
| "extract": { |
| "jsonPath": "$.task" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/question", |
| "name": "question", |
| "description": "Natural-language multiple-choice question.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "metadata_jsonl" |
| }, |
| "extract": { |
| "jsonPath": "$.question" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/options", |
| "name": "options", |
| "description": "List of multiple-choice answer options.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "metadata_jsonl" |
| }, |
| "extract": { |
| "jsonPath": "$.options" |
| } |
| }, |
| "repeated": true |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/answer", |
| "name": "answer", |
| "description": "Correct answer option as text.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "metadata_jsonl" |
| }, |
| "extract": { |
| "jsonPath": "$.answer" |
| } |
| } |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/videos", |
| "name": "videos", |
| "description": "List of video file paths associated with the question.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "metadata_jsonl" |
| }, |
| "extract": { |
| "jsonPath": "$.videos" |
| } |
| }, |
| "repeated": true |
| }, |
| { |
| "@type": "cr:Field", |
| "@id": "default/video_ranges", |
| "name": "video_ranges", |
| "description": "Optional temporal ranges or clip spans associated with the referenced videos; may be empty.", |
| "dataType": "sc:Text", |
| "source": { |
| "fileSet": { |
| "@id": "metadata_jsonl" |
| }, |
| "extract": { |
| "jsonPath": "$.video_ranges" |
| } |
| }, |
| "repeated": true |
| } |
| ] |
| } |
| ], |
| "rai:dataLimitations": "The simulation engines (CLEVRER, Kubric, Habitat) are inherently limited in visual diversity and photorealism compared to unconstrained real-world footage. The benchmark focuses strictly on visual, temporal, and spatial reasoning, explicitly excluding audio streams. The evaluation methodology is restricted to exact-match multiple-choice Question Answering (QA). Explicitly not recommended for: Evaluating models on open-ended conversational generation , or assessing real-world multimodal contexts where audio synchronization is critical.", |
| "rai:dataBiases": "Because the dataset is entirely procedurally generated via physics simulations and 3D mesh navigation, it contains no human subjects, eliminating demographic bias. \n\nDomain restrictions are present: Kubric and CLEVRER rely on abstract, frictionless geometric primitives (e.g., cubes, spheres) , while Habitat relies on the HM3D dataset, which intrinsically biases the topological data toward the specific indoor, architectural styles present in those original real-estate scans.", |
| "rai:personalSensitiveInformation": "None. The dataset relies entirely on open-source simulation engines and procedurally generated abstract environments, posing zero risk of containing personal or sensitive human data.", |
| "rai:dataUseCases": "Construct Validity: Measures cross-video reasoning in Multimodal Large Language Models (MLLMs) across four specific cognitive pillars: Temporal Alignment, Spatial Tracking, Comparative Reasoning, and Holistic Synthesis. \n\n\nEstablished Use Cases: Designed as a controlled, zero-shot diagnostic testbed for isolating reasoning failures and assessing spatial-temporal intelligence. \n\n\nNon-Established Use Cases: It is not intended to act as a direct proxy for evaluating models on noisy, unconstrained real-world video due to the sim-to-real visual gap.", |
| "rai:dataSocialImpact": "Positive Effects: Facilitates the development of more physically grounded and robust multimodal systems , which is necessary for downstream real-world applications like autonomous vehicle navigation or multi-camera security analysis. Negative Effects / Mitigations: As a foundational benchmark utilizing simulated geometries, it does not pose immediate negative societal impacts and inherently lacks a high risk for dual-use or malicious misuse.", |
| "rai:hasSyntheticData": true, |
| "prov:wasDerivedFrom": [ |
| { |
| "@id": "https://aihabitat.org/datasets/hm3d/", |
| "prov:label": "Habitat-Matterport 3D (HM3D)", |
| "sc:license": "Matterport End User License Agreement for Academic Use of Model Data", |
| "prov:wasAttributedTo": { |
| "@id": "https://aihabitat.org/", |
| "prov:label": "Meta AI & Matterport" |
| } |
| }, |
| { |
| "@id": "http://clevrer.csail.mit.edu/", |
| "prov:label": "CLEVRER (CoLlision Events for Video REpresentation and Reasoning)", |
| "sc:license": "MIT License / CC0 1.0 Universal", |
| "prov:wasAttributedTo": { |
| "@id": "https://www.csail.mit.edu/", |
| "prov:label": "MIT CSAIL, IBM Research, & DeepMind" |
| } |
| }, |
| { |
| "@id": "https://github.com/google-research/kubric", |
| "prov:label": "Kubric", |
| "sc:license": "Apache License 2.0", |
| "prov:wasAttributedTo": { |
| "@id": "https://research.google/", |
| "prov:label": "Google Research" |
| } |
| } |
| ], |
| "prov:wasGeneratedBy": [ |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q4929239" |
| }, |
| "prov:label": "Synthetic Video Generation", |
| "sc:description": "Generated synchronized multi-view video streams and extracted precise state matrices. Habitat was used to calculate topological trajectories via NavMesh and render them at 5 FPS with panoramic sweeps. Kubric simulated dynamic physical interactions of geometric primitives at 12 FPS using PyBullet. For the CLEVRER scenes, existing multi-object collision videos were utilized, parsing the comprehensive frame-by-frame state annotations to extract absolute physical ground truths and continuous kinematic trajectories.", |
| "prov:wasAttributedTo": [ |
| { |
| "@type": "prov:SoftwareAgent", |
| "@id": "habitat_simulator_engine", |
| "prov:label": "Habitat Simulator Engine", |
| "sc:description": "Calculated topological trajectories via NavMesh and rendered synchronized RGB, depth, and semantic observations at 5 FPS with panoramic sweeps." |
| }, |
| { |
| "@type": "prov:SoftwareAgent", |
| "@id": "kubric_generation_pipeline", |
| "prov:label": "Kubric Generation Pipeline", |
| "sc:description": "Simulated dynamic physical interactions of KuBasic geometric primitives at 12 FPS using the PyBullet physics engine and rendered via Blender Cycles." |
| }, |
| { |
| "@type": "prov:SoftwareAgent", |
| "@id": "clevrer_dataset_parser", |
| "prov:label": "CLEVRER Dataset Parser", |
| "sc:description": "Parsed existing comprehensive frame-by-frame state annotations from the CLEVRER corpus to extract absolute physical ground truths and continuous kinematic trajectories." |
| } |
| ] |
| }, |
| { |
| "@type": "prov:Activity", |
| "prov:type": { |
| "@id": "https://www.wikidata.org/wiki/Q109719325" |
| }, |
| "prov:label": "Programmatic Ground Truth & Distractor Synthesis", |
| "sc:description": "Extracted pixel-perfect ground truth programmatically directly from the underlying physics and rendering matrices of the simulators. No human annotators were utilized. To ensure robust evaluation of the models, deterministic logic generated mathematically hard distractors—such as the lowest Hamming distance permutations for temporal sequencing or exact secondary peak velocities for kinematic comparisons—to prevent heuristic guessing and minimize subjective ambiguity.", |
| "prov:wasAttributedTo": [ |
| { |
| "@type": "prov:SoftwareAgent", |
| "@id": "programmatic_generation_scripts", |
| "prov:label": "Programmatic Generation Scripts", |
| "sc:description": "Programmatically extracted pixel-perfect ground truth directly from simulation states and generated deterministic, mathematically hard distractors without human annotation." |
| } |
| ] |
| } |
| ], |
| "keywords": [ |
| "cross-video reasoning", |
| "multimodal evaluation", |
| "synthetic benchmark", |
| "video question answering", |
| "programmatic grounding" |
| ], |
| "isAccessibleForFree": true |
| } |
|
|
