{ "@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", "conformsTo": "http://mlcommons.org/croissant/1.0", "name": "SimVerse", "description": "A multi-task benchmark for evaluating multimodal LLMs on interactive simulation puzzles. Five independent tasks (Text-VOI placements, cube reconstruction, cube goal-roll, mechanical-arm lamp targeting, Cut the Rope video-to-command) share a uniform 9-section prompt skeleton and a uniform FINAL_JSON output contract for cross-task comparability. Each per-level record carries the literal prompt text the benchmark presents to models, so reproductions need no auxiliary code. Anonymized for double-blind review; author and citation fields contain placeholder values until the review concludes.", "license": "https://opensource.org/licenses/MIT", "url": "https://huggingface.co/datasets/SimVer-ano/simverse2026", "version": "1.0.0", "datePublished": "2026", "keywords": [ "benchmark", "multimodal", "spatial-reasoning", "video-understanding", "tool-use", "evaluation" ], "rai:dataCollection": "All five tasks were programmatically generated. VOI uses a polygon-rasterization XOR generator. cube1 and cube2 use a pure-Python cube-state simulator that records roll sequences and bottom-face imprints. lamp uses a forward-kinematics generator with axis-aligned obstacle placement. cutRope is built on top of the open-source yell0wsuit/cuttherope-h5dx HTML5 port (MIT-licensed) by recording deterministic gameplay clips and authoring matching command scripts. No human subjects or real-world data were involved.", "rai:dataAnnotationProtocol": "Reference solutions are produced by the same generators that create each puzzle (programmatic ground truth). For closed-form tasks (VOI, cube1, lamp), the answer is uniquely determined and machine-verifiable. For open-ended tasks (cube2 goal-roll, cutRope), the dataset's `answer` field carries one known-valid reference solution; validators run the underlying engine against the model's actual output rather than performing string equality, so multiple correct answers earn full credit.", "rai:dataAnnotationAnalysis": "Each generator's output was post-validated by the same engine the eval validator uses, so every shipped record is provably solvable. Sentinel '?' values appear in cube1 records when puzzle constraints leave a face under-determined; this is intentional and follows the documented sentinel-pair rule (rotation forced to 0 when patternId is '?').", "rai:dataPreprocessingProtocol": "All per-level JSON files conform to a v1 schema documented in PROMPT_SKELETON.md (locked schemas per task). A migration step lifts legacy task-native answer formats into the unified `answer` envelope and a separate populate_prompts step embeds the literal system+user prompt strings into each record. Both transformations are idempotent and reproducible from the accompanying code.", "rai:dataLimitations": "Synthetic puzzles in fixed visual styles per task — models may learn render-style shortcuts rather than the underlying reasoning skill. All prompts are English-only. Difficulty distribution is hand-tuned per generator and not necessarily uniform. cube1 includes '?' sentinel patterns when faces are under-determined; downstream uses outside the SimVerse evaluation flow may need to filter these. Because reference solutions for open-ended tasks are non-unique, simple string-match scoring is inappropriate and the bundled engine-based validator should be used.", "rai:dataBiases": "Each task's visual style is uniform across all of its records, which can bias evaluations toward render-style recognition rather than the reasoning skill the task is intended to probe. Reference solutions for open-ended tasks favor specific solution paths even when other paths are equally valid; validators are designed to score by simulation outcome rather than reference matching, but downstream comparisons that ignore this design will systematically under-credit divergent strategies. Object-count distributions in cutRope reflect the original yell0wsuit corpus and are not balanced across gameplay-element categories.", "rai:personalSensitiveInformation": "None. Fully synthetic; the dataset contains no personally identifiable information, no human-subject data, and no sensitive content. cutRope videos are gameplay recordings of an open-source physics-puzzle game with no real-world imagery.", "rai:dataUseCases": "Evaluating multimodal large language models on (a) spatial-pattern reconstruction (VOI, cube1, cube2), (b) multi-step planning under physical constraints (lamp, cube2, cutRope), and (c) short-horizon video-to-program inference (cutRope). The dataset is designed strictly for held-out benchmarking — there is no train split. It is NOT intended as a held-out training set, NOT a real-world capability predictor, and NOT a substitute for application-specific evaluation.", "rai:dataSocialImpact": "Low. Synthetic abstract puzzles are unlikely to encode harmful content or to enable misuse beyond benchmark gaming. Primary risk: if widely adopted, models tuned specifically on these visual styles may show inflated capability claims that do not transfer to real-world spatial reasoning tasks. The dataset is purely for evaluation and contains no instructions toward harmful behaviors.", "rai:dataReleaseMaintenancePlan": "All levels are deterministically regenerable from the generators in the accompanying code repository (linked from the dataset card under the post-acceptance camera-ready URL). Schema and prompt-template changes are released as new versioned snapshots; per-record `legacy_*` fields preserve pre-migration values for one release cycle to ease comparisons across versions.", "distribution": [ { "@type": "cr:FileObject", "@id": "voi-test-jsonl", "name": "voi/test.jsonl", "description": "Per-record JSON-Lines for the VOI (Text-VOI placements) task, one level per line.", "contentUrl": "voi/test.jsonl", "encodingFormat": "application/jsonlines", "sha256": "cfd99ae4cf52689d26c6e7d2488a124636c2d8aa73d3f48ed80cd916b8edb572", "contentSize": "4230614 B" }, { "@type": "cr:FileSet", "@id": "voi-images", "name": "voi/images", "description": "Per-level rendered images (target pattern + base-shape pieces) for VOI.", "encodingFormat": "image/png", "includes": "voi/images/**/*.png" }, { "@type": "cr:FileObject", "@id": "cube1-test-jsonl", "name": "cube1/test.jsonl", "description": "Per-record JSON-Lines for the cube reconstruction task.", "contentUrl": "cube1/test.jsonl", "encodingFormat": "application/jsonlines", "sha256": "40d836f859c2ae775ba7649702623ea2f7d48410d8c394ee78feda06d42dfb3a", "contentSize": "4680065 B" }, { "@type": "cr:FileSet", "@id": "cube1-images", "name": "cube1/images", "description": "Blank-cross-net and path-imprint images for cube reconstruction.", "encodingFormat": "image/png", "includes": "cube1/images/**/*.png" }, { "@type": "cr:FileObject", "@id": "cube2-test-jsonl", "name": "cube2/test.jsonl", "description": "Per-record JSON-Lines for the cube goal-roll task.", "contentUrl": "cube2/test.jsonl", "encodingFormat": "application/jsonlines", "sha256": "3055027fd6946cc6711be3bc6717347b80266e80c2fd8dde48844df9f9e3413f", "contentSize": "4525952 B" }, { "@type": "cr:FileSet", "@id": "cube2-images", "name": "cube2/images", "description": "Initial-net and target-top-face images for cube goal-roll.", "encodingFormat": "image/png", "includes": "cube2/images/**/*.png" }, { "@type": "cr:FileObject", "@id": "lamp-test-jsonl", "name": "lamp/test.jsonl", "description": "Per-record JSON-Lines for the mechanical-lamp targeting task.", "contentUrl": "lamp/test.jsonl", "encodingFormat": "application/jsonlines", "sha256": "6d92c608683f02228ca5dc32c448863889a1aed441960b9df18dac1dbeb77889", "contentSize": "3742112 B" }, { "@type": "cr:FileSet", "@id": "lamp-images", "name": "lamp/images", "description": "Workspace images (arm, target, obstacles) for the mechanical-lamp task.", "encodingFormat": "image/png", "includes": "lamp/images/**/*.png" }, { "@type": "cr:FileObject", "@id": "cutrope-test-jsonl", "name": "cutrope/test.jsonl", "description": "Per-record JSON-Lines for the Cut the Rope video-to-command task.", "contentUrl": "cutrope/test.jsonl", "encodingFormat": "application/jsonlines", "sha256": "06b28abdf1d57be23abc18c7e57211bb8e5efcc017e746c7b3a9268c9ed35235", "contentSize": "2634596 B" }, { "@type": "cr:FileSet", "@id": "cutrope-videos", "name": "cutrope/videos", "description": "Short MP4 gameplay clips (~3s each, 1920x1080) for cutRope.", "encodingFormat": "video/mp4", "includes": "cutrope/videos/*.mp4" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "voi", "name": "voi", "description": "Text-VOI placements task: 600 records.", "field": [ { "@type": "cr:Field", "@id": "voi/sample_id", "name": "sample_id", "description": "Stable level id, e.g. voi-000.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "voi-test-jsonl" }, "extract": { "jsonPath": "$.__sample_id__" } } }, { "@type": "cr:Field", "@id": "voi/prompt_system", "name": "prompt_system", "description": "Verbatim system-prompt text the benchmark presents.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "voi-test-jsonl" }, "extract": { "jsonPath": "$.prompt.system" } } }, { "@type": "cr:Field", "@id": "voi/prompt_user", "name": "prompt_user", "description": "Verbatim user-prompt text (9-section skeleton, instance-filled).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "voi-test-jsonl" }, "extract": { "jsonPath": "$.prompt.user" } } }, { "@type": "cr:Field", "@id": "voi/answer", "name": "answer", "description": "Reference placements list. Schema: {placements:[{shape,angle,vertex,grid}]}. (Stored as a JSON-encoded string.)", "dataType": "sc:Text", "source": { "fileObject": { "@id": "voi-test-jsonl" }, "extract": { "jsonPath": "$.answer" } } } ] }, { "@type": "cr:RecordSet", "@id": "cube1", "name": "cube1", "description": "Cube reconstruction task: 502 records.", "field": [ { "@type": "cr:Field", "@id": "cube1/sample_id", "name": "sample_id", "description": "Stable level id, e.g. C001.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube1-test-jsonl" }, "extract": { "jsonPath": "$.__sample_id__" } } }, { "@type": "cr:Field", "@id": "cube1/prompt_system", "name": "prompt_system", "description": "Verbatim system-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube1-test-jsonl" }, "extract": { "jsonPath": "$.prompt.system" } } }, { "@type": "cr:Field", "@id": "cube1/prompt_user", "name": "prompt_user", "description": "Verbatim user-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube1-test-jsonl" }, "extract": { "jsonPath": "$.prompt.user" } } }, { "@type": "cr:Field", "@id": "cube1/answer", "name": "answer", "description": "Reference six-face map. Schema: {faces:{TOP:{patternId,rotation},...}}. (Stored as a JSON-encoded string.)", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube1-test-jsonl" }, "extract": { "jsonPath": "$.answer" } } } ] }, { "@type": "cr:RecordSet", "@id": "cube2", "name": "cube2", "description": "Cube goal-roll task: 502 records.", "field": [ { "@type": "cr:Field", "@id": "cube2/sample_id", "name": "sample_id", "description": "Stable level id, e.g. C001.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube2-test-jsonl" }, "extract": { "jsonPath": "$.__sample_id__" } } }, { "@type": "cr:Field", "@id": "cube2/prompt_system", "name": "prompt_system", "description": "Verbatim system-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube2-test-jsonl" }, "extract": { "jsonPath": "$.prompt.system" } } }, { "@type": "cr:Field", "@id": "cube2/prompt_user", "name": "prompt_user", "description": "Verbatim user-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube2-test-jsonl" }, "extract": { "jsonPath": "$.prompt.user" } } }, { "@type": "cr:Field", "@id": "cube2/answer", "name": "answer", "description": "One known-valid roll sequence. Schema: {directions:['N'|'S'|'E'|'W',...]}. (Stored as a JSON-encoded string.)", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cube2-test-jsonl" }, "extract": { "jsonPath": "$.answer" } } } ] }, { "@type": "cr:RecordSet", "@id": "lamp", "name": "lamp", "description": "Mechanical-lamp targeting task: 610 records.", "field": [ { "@type": "cr:Field", "@id": "lamp/sample_id", "name": "sample_id", "description": "Stable level id, e.g. lamp-000.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "lamp-test-jsonl" }, "extract": { "jsonPath": "$.__sample_id__" } } }, { "@type": "cr:Field", "@id": "lamp/prompt_system", "name": "prompt_system", "description": "Verbatim system-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "lamp-test-jsonl" }, "extract": { "jsonPath": "$.prompt.system" } } }, { "@type": "cr:Field", "@id": "lamp/prompt_user", "name": "prompt_user", "description": "Verbatim user-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "lamp-test-jsonl" }, "extract": { "jsonPath": "$.prompt.user" } } }, { "@type": "cr:Field", "@id": "lamp/answer", "name": "answer", "description": "Reference per-joint angles. Schema: {actions:[{joint,angle},...]}. (Stored as a JSON-encoded string.)", "dataType": "sc:Text", "source": { "fileObject": { "@id": "lamp-test-jsonl" }, "extract": { "jsonPath": "$.answer" } } } ] }, { "@type": "cr:RecordSet", "@id": "cutrope", "name": "cutrope", "description": "Cut the Rope video-to-command task: 272 records.", "field": [ { "@type": "cr:Field", "@id": "cutrope/sample_id", "name": "sample_id", "description": "Stable level id, e.g. rope-000.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cutrope-test-jsonl" }, "extract": { "jsonPath": "$.__sample_id__" } } }, { "@type": "cr:Field", "@id": "cutrope/prompt_system", "name": "prompt_system", "description": "Verbatim system-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cutrope-test-jsonl" }, "extract": { "jsonPath": "$.prompt.system" } } }, { "@type": "cr:Field", "@id": "cutrope/prompt_user", "name": "prompt_user", "description": "Verbatim user-prompt text.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cutrope-test-jsonl" }, "extract": { "jsonPath": "$.prompt.user" } } }, { "@type": "cr:Field", "@id": "cutrope/answer", "name": "answer", "description": "Reference 3-star command script. Schema: {commands,reason,confidence}. (Stored as a JSON-encoded string.)", "dataType": "sc:Text", "source": { "fileObject": { "@id": "cutrope-test-jsonl" }, "extract": { "jsonPath": "$.answer" } } } ] } ], "citeAs": "@dataset{simverse_anonymous_2026,\n title = {SimVerse: A Multi-Task Benchmark for Multimodal Reasoning on Interactive Simulation Puzzles},\n author = {Anonymous Authors (under double-blind review)},\n year = {2026},\n url = {https://huggingface.co/datasets/SimVer-ano/simverse2026}\n}", "rai:hasSyntheticData": true, "rai:hasPersonalInformation": false, "rai:hasSensitivePersonalInformation": false }