Live Public Release: spacetime-ml/cad-cua • 401 Verified Files • 15.48 GB • CC BY 4.0

spacetime-ml/cad-cua: 3D CAD CUA & Deterministic B-Rep RLVR

The public multimodal 3D Computer-Use Agent (CUA) & Reinforcement Learning with Verifiable Rewards (RLVR) dataset and evaluation suite—pairing 62 commercial CAD episodes, 48 ISO-128 2D Reference Engineering Blueprints, and 215 mechanical CAD deliverables with verify_step.py deterministic ISO-10303-21 .STEP grading.

401 Files

15.48 GB Public HF Repo

SHA256SUMS-verified · CC BY 4.0

62 Episodes

6 Commercial/Open CAD Suites

34,551 PNGs · 260,203 GUI events

215 CAD Files

48 Submission Archives

115 native + 98 STEP + 2 mesh

85.0%+ Fail

20-Task .STEP RLVR Benchmark

Claude 4.5 (15%), GPT-5.2 & Gemini 3 (10%)

The 2D-Blueprint-to-3D B-Rep Spatial Reasoning Bottleneck

Frontier multimodal models achieve high scores on 2D browser clicking and static visual QA, but fail 85.0% to 90.0% of the time when reconstructing watertight ISO-10303-21 .STEP solids from 4-view ISO-128 engineering blueprints—suffering from missing internal counterbores, fillet self-intersections, and non-manifold open shells.

View 20-Task Leaderboard
spacetime-ml/cad-cua•401 Public Files (15.48 GB) • CC BY 4.0

CAD-CUA: 62 Industrial 3D Trajectories, 48 Blueprints & 215 CAD Deliverables

Hosted on Hugging Face Hub across 3 configurations (default, models, and benchmark), spacetime-ml/cad-cua pairs 62 verified CAD sessions across SolidWorks (51), PTC Creo (9), AutoCAD (1), FreeCAD (1), Autodesk Fusion (1), and Autodesk Inventor (1) with 48 ISO-128 2D Reference Engineering Blueprints, 215 mechanical CAD deliverables (115 native + 98 STEP + 2 mesh), and verify_step.py.

62 Episodes

48 Model + 14 Trajectory

34,551 masked PNG frames (72,070 raw inspected; 9-pass OCR with 0 errors)

48 Blueprints

ISO-128 2D Reference Pairs

4-view Third-Angle orthographic + isometric drawings (.png & .pdf)

215 CAD Files

115 Native + 98 STEP + 2 Mesh

48 models/cad-XXXXXX.zip archives 100% paired with episodes

62 Previews

PNG + H.264 MP4 + PDF Dossiers

62 keyframes, 62 cursor-annotated MP4 videos, and 62 1-page storyboard PDFs

Live RLVR Verification Benchmark • data/benchmark.jsonl

20-Task Frontier Model RLVR Benchmark (Watertight Manifold CLOSED_SHELL ≥ 1, OPEN_SHELL == 0 & R ≥ 0.85)

python3 verify_step.py --candidate pred.step --ground-truth gt.step
Model / AgentPass@1 (20 Blueprints)Failure RateMean Reward (R)BBox IoUTopology ScorePrimary Failure Mode
Human CAD Engineer (Ground Truth)100.0% (20/20)0.0%1.0001.0001.000Watertight ISO-10303-21 B-Rep manifold solid
Claude 4.5 (Computer Use / CAD Code)15.0% (3/20)85.0% (17/20)0.4630.5430.415Missing internal counterbores, fillet self-intersections
GPT-5.2 (Vision + CAD Code)10.0% (2/20)90.0% (18/20)0.3910.4760.3353D axis swap, non-manifold shells, bore pitch drift
Gemini 3 (Multimodal Spatial + CAD)10.0% (2/20)90.0% (18/20)0.3710.4490.324Unclosed sketch loops, missing radial hole arrays
Live Dataset Sample Explorer · spacetime-ml/cad-cua

Inspect Real 2D Blueprints, CUA GUI Trajectories & B-Rep Rewards

1. 2D ISO-128 Reference Blueprint (cad-000020.png)cad-bench-06
2D Engineering Blueprint for cad-000020

Construct a watertight 3D B-Rep solid matching 2D Reference Blueprint CAD-000020 (outer envelope 100.00 × 67.50 × 35.00 mm, 78 B-Rep faces: 14 planar, 38 cylindrical) within ISO 2768-mK tolerances.

Envelope (mm)
100×67.5×35
B-Rep Topology
78F / 170E / 100V
Volume (mm³)
200,812.5
2. CUA GUI Trajectory (SolidWorks)
Deliverable STEP:deliverables/AX-001.STEP
Retained Frames / Events:413 frames · 9,986 events
Session Span / Canvas:37.8 min · 1366×768
9-Pass OCR Audited (0 PII)Download cad-000020.zip ↗
3. B-Rep RLVR Verifierverify_step.py
Human Reference (.STEP)PASS · R=1.000
mode: verified_pass (χ=2)
Claude 4.5 SonnetPASS · R=0.975
mode: verified_pass
GPT-5.2 ThinkingPASS · R=0.975
mode: verified_pass
Gemini 3 ProFAIL · R=0.453
mode: dimensional_tolerance_violation
Reproduce Locally (<50ms):
python3 verify_step.py \
  --ground-truth models/cad-000020.zip \
  --candidate models/cad-000020.zip
Interactive In-Browser ISO-10303-21 .STEP B-Rep Verifier Sandbox

Test verify_step.py Live in Your Browser (<10ms Execution)

Select a reference blueprint below, toggle between Human Ground Truth and Frontier Model candidate .STEP files—or drag-and-drop your own .STEP / .STP file to compute exact B-Rep RLVR rewards in real time.

Load Candidate .STEP:
1. Candidate ISO-10303-21 .STEP Buffer (Editable)cad-000021-human-gt.step

Tip: Try editing CLOSED_SHELL to OPEN_SHELL or modifying CARTESIAN_POINT coordinates above to watch the B-Rep reward update instantaneously.

Parse Latency: 0.6 ms0 3D Cartesian Points
2. Extracted B-Rep Invariants (Cand vs. GT)χ = 0
Manifold ClosureGT: CLOSED_SHELL (OK)
Cand: OPEN_SHELL (0)
Principal BBox (mm)GT: 0×0×0
Cand: 0×0×0
B-Rep Faces / EdgesGT: 0F / 0E / 0V
Cand: 0F / 0E / 0V
Analytical SurfacesGT: 0 Plane · 0 Cyl
Cand: 0 Plane · 0 Cyl
Est. Solid VolumeGT: 0 mm³
Cand: 0 mm³
Bore Radii Spectrum: None detected
3. RLVR Reward OutputREJECTED / FAIL
Composite RLVR Reward (R)
0.0000
invalid_step_syntax
s_bbox (35% weight)0.0000
s_vol_area (25% weight)0.0000
s_topology (20% weight)0.0000
s_surface (20% weight)0.0000
Formula: R = 0.35·s_bbox + 0.25·s_vol + 0.20·s_top + 0.20·s_surf
Zero-Contamination Private Holdout & Multi-Engineer DPO Vault

Held-Out Enterprise Split: 105 Verified Sessions & 29 Multi-Trajectory Part Families

While 62 episodes are published openly in spacetime-ml/cad-cua, SpacetimeML maintains an air-gapped, continuously expanding private holdout (currently 105 verified human CAD sessions · 72.12 hours · 102,821 frames · 502,556 OS events · 0 duplicate deliverable SHA-256 hashes) featuring 29 multi-engineer part families (part_family_id) where 2–4 distinct mechanical engineers independently model the exact same 2D blueprint.

Total Verified Pool
105 Sessions
62 Public HF + 43 Air-Gapped Holdout (72.12 hrs)
Full Telemetry Corpus
502,556 Events
102,821 frames · 100% dynamic app-context verified
Deliverable Uniqueness
0 Duplicates
Cryptographic SHA-256 deduplicated across all 105 sessions
DPO / Pass@k Families
29 Part Families
2–4 independent human trajectories per blueprint
Sample Multi-Engineer Part Families (part_family_id) — Paired Human Trajectories for DPO & GRPO
All trajectories converge to verified ISO-10303-21 .STEP solids (R = 1.000)
Part Family IDBlueprint Spec & ComponentEngineersSession Durations (Min)Horizon SpreadFrames / EventsPost-Training RL / DPO Signal
pf_model_mania_2011[AX-145 / 018]Model Mania 2011 Ribbed Clevis & Boss4× Human10.4m · 13.6m · 25.2m · 46.0m4.4× efficiency spread2,309f / 15,316eDPO Action-Horizon Pruning + Pass@4 GRPO
pf_ax_167[AX-167 / 174]Multi-Port Hydraulic Valve Block4× Human15.3m · 18.7m · 32.6m · 64.2m4.2× efficiency spread3,748f / 5,909eFeature-Tree Ordering Preference Pairs
pf_model_mania_2008[AX-023 / 119]Model Mania 2008 Swept Bellcrank Yoke3× Human20.8m · 68.1m · 101.6m4.9× efficiency spread3,815f / 42,647eCross-Suite Trajectory Alignment (SolidWorks / Creo)
pf_handle_phase1[AX-123 / 106]Ergonomic Contoured Actuator Handle3× Human31.2m · 80.1m · 95.7m3.1× efficiency spread5,204f / 15,229eSurface Fillet & Shell Recovery Trajectories
pf_model_mania_2014[AX-118 / 126]Model Mania 2014 Tri-Boss Trunnion3× Human29.5m · 49.6m · 68.1m2.3× efficiency spread4,331f / 4,364eMulti-Datum Sketch Plane Grounding
pf_ax_166[AX-166 / 176]Slotted Eccentric Cam Housing3× Human21.4m · 37.6m · 60.6m2.8× efficiency spread3,245f / 3,504eConcentric Bore Constraint Satisfaction
pf_ax_105[AX-105]Counterbored Bearing Pillow Block3× Human28.9m · 36.5m · 39.9m1.4× efficiency spread2,839f / 2,849eConsistent B-Rep Topology Convergence (χ=2)
pf_stepped_shaft[AX-017 / 014]Keyed Multi-Diameter Drive Shaft3× Human8.2m · 14.1m · 17.2m2.1× efficiency spread797f / 955eRevolve vs. Stacked Extrude DPO Pairs
Dataset Configurations & Modalities

Inside the 401-File Public Release

Every file in spacetime-ml/cad-cua is SHA256SUMS-verified, OCR-scrubbed for privacy, and structured for immediate Hugging Face datasets loading.

62 Episodes · 6 Suites260,203 Events

default (data/train.jsonl) — 62 Multimodal CAD Episodes

62 retained CAD sessions (48 verified_cad_model + 14 verified_trajectory) across 6 CAD software suites: SolidWorks (51), PTC Creo (9), AutoCAD (1), FreeCAD (1), Autodesk Fusion (1), and Autodesk Inventor (1). Includes 34,551 privacy-masked PNG observations and 260,203 desktop telemetry events.

Repository: spacetime-ml/cad-cua
34,551 Masked PNGs
48 Paired Archives215 Deliverables

models (data/models.jsonl) — 48 Archives & 215 CAD Files

48 submission archives (models/cad-XXXXXX.zip) 100% paired with episode observations (0 orphan archives). Contains 215 mechanical CAD deliverables: 115 native parametric files (.sldprt, .sldasm, .prt, .asm, .fcstd, .ipt, .dwg, .f3d), 98 ISO-10303-21 .step/.stp B-Rep solids, and 2 mesh/exchange files.

Repository: spacetime-ml/cad-cua
115 Native + 98 STEP + 2 Mesh
PNG + Vector PDF48 Blueprint Pairs

48 ISO-128 2D Reference Engineering Blueprints

48 standardized engineering drawing pairs (blueprints/cad-XXXXXX.png & blueprints/cad-XXXXXX.pdf) featuring 4-view Third-Angle orthographic projections (Front, Top, Right) + Isometric wireframe views with explicit bounding dimensions and bore callouts.

Repository: spacetime-ml/cad-cua
ISO-128 4-View + Dim Callouts
Pass: Manifold + R ≥ 0.8585.0%+ Fail Rate

benchmark (data/benchmark.jsonl) — 20-Task Frontier RLVR Eval

20 held-out 2D-blueprint-to-3D-B-Rep reconstruction tasks evaluated with verify_step.py across principal-axis BBox IoU, closed-shell volume/area, Euler-Poincaré topology, and analytical cylindrical/conical/toroidal surface spectra.

Repository: spacetime-ml/cad-cua
Claude 4.5, GPT-5.2, Gemini 3
Deterministic B-Rep Reward Engine

Mathematical Verification with verify_step.py

How verify_step.py computes a continuous RLVR shaping reward R in [0, 1] directly from ISO-10303-21 .STEP files.

35% Principal-Axis Bounding Box IoU (s_bbox)

Computes rotation-invariant sorted 3D principal extents and PCA inertia moments from VERTEX_POINT and CARTESIAN_POINT coordinates, catching 3D axis swaps and dimensional tolerance violations.

25% Volume & Surface-Area Conservation (s_vol_area)

Evaluates solid packing density, 3D spatial occupancy, and closed-shell surface area conservation across MANIFOLD_SOLID_BREP and CLOSED_SHELL entities.

40% B-Rep Topology (20%) & Analytical Surfaces (20%)

Combines 20% Euler-Poincaré topological graph fidelity (faces, edges, vertices, CLOSED_SHELL ≥ 1, OPEN_SHELL == 0) with 20% analytical surface & cylindrical bore radius spectrum matching.

verify_step.py • Composite RLVR Reward Formula
Watertight Manifold (CLOSED_SHELL ≥ 1, OPEN_SHELL == 0) & R ≥ 0.85
{
  "episode_id": "cad-000001",
  "cad_application": "SolidWorks",
  "prompt_blueprint": "blueprints/cad-000001.png",
  "ground_truth_archive": "models/cad-000001.zip",
  "verifier": "verify_step.py",
  "reward_weights": {
    "s_bbox": 0.35,      // Principal-Axis Bounding Box IoU
    "s_vol_area": 0.25,  // Closed-Shell Volume & Surface-Area Conservation
    "s_topology": 0.20,  // B-Rep Topological Graph & Euler-Poincare Fidelity (V - E + F)
    "s_surface": 0.20    // Analytical Surface & Cylindrical Bore Radius Spectrum
  },
  "pass_criterion": {
    "closed_shells_min": 1,
    "open_shells_max": 0,
    "min_rlvr_reward": 0.85
  }
}
Private Enterprise CAD-RLVR Expansions

Need 50,000+ Held-Out 2D-to-3D B-Rep Tasks & Custom Human CAD Trajectories?

Beyond the public 62-episode spacetime-ml/cad-cua release, SpacetimeML provisions private, zero-contamination parametric CAD corpora, multi-part assembly benchmarks, and high-throughput verify_step.py RL rollout clusters for frontier AI labs.

50,000+ Procedural ISO-128 Blueprint + .STEP Tasks
Managed SolidWorks, Creo, NX & Inventor Engineers
Ray / PyTorch / GRPO B-Rep Reward Workers

Ready to benchmark and train your 3D CAD agents?

Download the public 401-file spacetime-ml/cad-cua dataset on Hugging Face or request private enterprise CAD-RLVR benchmark access.