{ "@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": "MUSE", "description": "MUSE (Multi-modal CAD Understanding & Synthesis Evaluation) is a benchmark of 106 engineering CAD design cases. Each case pairs a natural-language design specification with two complementary ground-truth artefacts (a 4-view dimensioned engineering drawing and a 3D STP rendered image) and a hand-crafted, rubric-style evaluation guide covering assembly readiness, joint design, tolerance, and functional adaptation. The benchmark targets evaluation of large language and multi-modal models on engineering-grade CAD generation and understanding.", "conformsTo": "http://mlcommons.org/croissant/1.0", "version": "1.0.0", "url": "https://huggingface.co/datasets/dongxiaoyu/MUSE", "license": "https://creativecommons.org/licenses/by/4.0/", "citeAs": "@inproceedings{muse_2026, title={MUSE: A CAD Design Benchmark with Multi-modal Ground Truth and Rubric-based Evaluation}, author={Dong, Xiaoyu and others}, booktitle={Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track}, year={2026}}", "creator": { "@type": "Person", "name": "Xiaoyu Dong", "url": "https://huggingface.co/dongxiaoyu" }, "keywords": [ "cad", "3d", "benchmark", "engineering design", "evaluation", "rubric", "multi-modal", "image-to-text", "text-to-image" ], "datePublished": "2026-05-07", "rai:dataCollection": "CAD geometry seeds were generated with assistance from leading commercial large language models (GPT-4 family, Claude, Gemini) between 2026-04 and 2026-05. For each case, the authors prompted the LLM to produce CadQuery Python code, executed the code in a sandboxed environment to obtain a STEP file, generated a 4-view dimensioned engineering drawing (SVG, then converted to PNG via rsvg-convert), and produced a 3D rendered PNG using a VTK-based off-screen renderer. Natural-language design specifications and per-case evaluation rubrics were authored entirely by the human authors. Every artefact was reviewed and curated by the authors before inclusion.", "rai:dataCollectionType": "Synthetic, author-curated. No human subjects, no scraped web content, no third-party copyrighted CAD models.", "rai:dataCollectionTimeFrame": [ { "@type": "sc:Date", "@value": "2026-04-01" }, { "@type": "sc:Date", "@value": "2026-05-07" } ], "rai:dataPreprocessingProtocol": "STEP files generated by CadQuery were validated for watertightness and self-intersection-freeness. SVG drawings were rendered to PNG with rsvg-convert at the SVG's native resolution. For 9 organic / curved-vase cases, the SVG projection cannot capture the surface geometry, so the engineering-drawing PNG and the 3D-render PNG are intentionally identical (the curved silhouette is used for both ground-truth views). For 6 cases where the upstream pipeline did not produce a thumbnail render, a fresh VTK render was produced from the STEP file at submission time.", "rai:dataAnnotationProtocol": "Per-case evaluation rubrics were authored by the team. Each rubric contains 4–8 independent criteria covering: (1) Assembly Readiness (component graph topology), (2) Joint Design (physical connection type), (3) Tolerance (manufacturing precision), (4) Functional Adaptation (intended-use fitness), and additional case-specific criteria. Each criterion is a binary 0/1 (Pass/Fail) judgement with an explicit Pass description and an explicit Fail description. Rubrics were drafted by referencing the corresponding ground-truth SVG and STEP render; ambiguity was resolved by collective review.", "rai:dataAnnotationPlatform": "Internal markdown editing; no third-party annotation platform.", "rai:dataAnnotationAnalysis": "Inter-rater reliability has not been formally measured because rubrics were authored by a single coordinated team rather than independent crowd-workers. Future versions will include held-out rubric checks.", "rai:dataAnnotatorDemographics": "Annotations (rubrics and design specifications) were authored by the paper's co-authors, all of whom are graduate-level researchers in mechanical engineering or computer science with prior CAD experience.", "rai:dataUseCases": "Primary use case: benchmarking large language models and vision-language models on their ability to (a) interpret natural-language CAD design specifications, (b) generate or reason about engineering CAD geometry, and (c) reason about manufacturability concepts including assembly topology, joint type selection, tolerance, and functional adaptation. Secondary use case: rubric-based qualitative analysis of model outputs against engineering-grade ground truth.", "rai:dataBiases": "1. Domain coverage is biased toward Western-style household objects and small fixtures (chairs, tables, shelves, vases, holders, pegboards, stools); the dataset does not cover automotive, aerospace, electronics, or large civil-engineering domains. 2. Manufacturing methods are limited to CNC milling, 3D printing, traditional joinery, and a few adhesive/fastening methods; there is no casting, forging, or additive metal manufacturing. 3. Material vocabulary is dominated by timber, plastic, and metal. 4. Component counts skew toward 10-60 parts; very large or very small assemblies are rare. 5. LLM-assisted seeding may inherit stylistic and naming biases of the underlying foundation models. 6. All design specifications are in English, so multilingual CAD comprehension cannot be evaluated.", "rai:dataLimitations": "1. Only 4-view orthographic + isometric drawings and a single 3D render are provided; no live STEP geometry comparison is included in the rubric scoring. 2. STEP files are watertight but tolerance/clearance values are not validated against physical fabrication; manufacturability is judged visually rather than by physical assembly. 3. Rubric scoring is binary (0/1) per criterion; partial credit is not encoded, which can over-penalize models that produce mostly-correct designs. 4. Each rubric is authored against one specific reference design and may be brittle when applied to substantially different but equally valid alternative designs. 5. With 106 cases, the benchmark provides moderate statistical resolution but cannot finely separate models with very small performance differences.", "rai:annotationsPerItem": "1 design description + 1 evaluation rubric (with 4-8 binary criteria) per case.", "rai:personalSensitiveInformation": "None. The dataset contains only synthetic CAD geometry, English-language design specifications written by the authors, and rubric text. No human-subject data, no images of people, no audio or video, no PII, no health, financial, location, or biometric information are included.", "rai:sensitiveData": "None of the categories tracked by the Croissant RAI specification (gender, race, ethnicity, religion, sexual orientation, political beliefs, financial data, health data, geographic location, biometric data, or genetic data) appear in the dataset.", "rai:dataSocialImpact": "Positive impact: rigorous evaluation of LLM/VLM ability on engineering tasks supports research into AI-assisted design tools that augment human designers, lower the barrier to entry for novice mechanical designers, and enable faster iteration on early-stage product design. Negative impact: improved model capabilities on engineering-grade CAD generation could be misused to reproduce proprietary mechanical designs without authorization or to bypass professional review in safety-critical applications. The dataset itself contains no operational hazards (no weapons, surveillance content, or dual-use components) and the synthetic-only collection method avoids privacy harms during dataset construction.", "rai:hasSyntheticData": true, "isLiveDataset": false, "distribution": [ { "@type": "cr:FileObject", "@id": "metadata-jsonl", "name": "metadata.jsonl", "description": "One JSON object per case (106 cases total). Each object points to the four files that make up the case: design description (input), 4-view engineering-drawing PNG (ground truth view 1), STP-rendered PNG (ground truth view 2), and the per-case evaluation rubric.", "contentUrl": "https://huggingface.co/datasets/dongxiaoyu/MUSE/resolve/main/metadata.jsonl", "encodingFormat": "application/jsonlines", "sha256": "a2b7ac9453b3bedde6f5fd65748e5fd5258fff209896cc28da4b5c1514a83868" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "cases", "name": "cases", "description": "One record per benchmark case (106 total). Each record links the design description, the two ground-truth images, and the evaluation rubric. Records are streamed from metadata.jsonl.", "key": { "@id": "cases/case_id" }, "field": [ { "@type": "cr:Field", "@id": "cases/case_id", "name": "case_id", "description": "Unique identifier of the case (e.g. 'bookshelf', 'chair_2', 'vase_wave_blossom').", "dataType": "sc:Text", "source": { "fileObject": { "@id": "metadata-jsonl" }, "extract": { "column": "case_id" } } }, { "@type": "cr:Field", "@id": "cases/design_description", "name": "design_description", "description": "Repository-relative path to the case's design description markdown file (input prompt).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "metadata-jsonl" }, "extract": { "column": "design_description" } } }, { "@type": "cr:Field", "@id": "cases/svg_png", "name": "svg_png", "description": "Repository-relative path to the 4-view dimensioned engineering drawing PNG (ground truth, view #1).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "metadata-jsonl" }, "extract": { "column": "svg_png" } } }, { "@type": "cr:Field", "@id": "cases/stp_render", "name": "stp_render", "description": "Repository-relative path to the 3D STP-rendered PNG (ground truth, view #2). For 9 organic/curved-vase cases this is intentionally identical to svg_png.", "dataType": "sc:Text", "source": { "fileObject": { "@id": "metadata-jsonl" }, "extract": { "column": "stp_render" } } }, { "@type": "cr:Field", "@id": "cases/evaluation_rubric", "name": "evaluation_rubric", "description": "Repository-relative path to the per-case evaluation rubric markdown file (4-8 binary criteria).", "dataType": "sc:Text", "source": { "fileObject": { "@id": "metadata-jsonl" }, "extract": { "column": "evaluation_rubric" } } } ] } ] }