Datasets:

CAD-bench
/

cad-bench-ed-2026-anonymous

text stringlengths 0 132
Create a simple aluminum-extrusion-like bar segment.

Requirements:
- Outer body is a 20.0 x 20.0 mm square prism, length 60.0 mm.
- Axis is global Z.
- Bottom face lies on z=0 and the part is entirely in z>=0.
- Through-bores run parallel to +Z over full length:
- one center bore: diameter 5.0 mm at (x, y) = (0, 0)
- four bores: diameter 3.0 mm at (±7.0, 0) and (0, ±7.0)
Create one rectangular box.

Requirements:
- Dimensions: 30.0 mm (X) x 20.0 mm (Y) x 10.0 mm (Z).
- Faces aligned with global axes.
- Bottom face lies on z=0 and the part is entirely in z>=0.
- Center on X/Y is at (0, 0).
Create a three-shaft idler gearbox on three pre-existing D-shafts.

Rig (already present; do not model):
- Input shaft center at (0, 0), idler shaft center at (20, 0), output shaft center at (50, 0)
- All shafts are along +Z
- All shafts are D-shafts: nominal diameter 4 mm with flat at x = center_x + 1.5 mm
- Shaft z-range is [0, 10]
- Input speed +120 rpm

Your part requirements:
- Include only transmission geometry in z in [0, 10]
- Use three rigid bodies:
- one input gear body on the input shaft
- one idler gear body on the middle shaft
- one output gear body on the output shaft
- A direct three-gear train is expected
- Target output speed is about +55 rpm
- Evaluation uses simultaneous rigid-body physical simulation on your authored
mesh geometry directly, on the fixed D-shafts above, at the stated +120 rpm input speed

Scoring points:
- correct physical transfer ratio and direction under simulation
- preserved authored placement on all three shafts
- correct three-body gear-train structure
Create a compound right-angle gearbox that reverses output direction.

Hidden evaluation mounts the model onto three fixed perpendicular shafts and runs a rigid-body physical simulation.
Model only the gears, not the shafts or housing.

Shaft layout:
- Input shaft: global Z axis through (x, y) = (12, 0)
- Compound shaft: global X axis through (y, z) = (0, 18)
- Output shaft: global Y axis through (x, z) = (40, 18)

Gear requirements:
- Four separate gears total:
- one gear on the input shaft
- two gears on the compound shaft
- one gear on the output shaft
- Keep the whole model inside:
- x in [4, 48]
- y in [-11, 11]
- z in [7, 29]
- Approximate outer diameters:
- input bevel gear: 15 mm
- compound stage-1 bevel gear: 21 mm
- compound stage-2 bevel gear: 11 mm
- output bevel gear: 15 mm
- Use a normal bevel-gear style layout with matched pitch-cone geometry for both right-angle stages
- The compound shaft should carry two distinct gears centered near x=17 and x=35
- Make the first right-angle stage visibly larger than the second stage
- The authored gear bodies must be disjoint at rest: zero body-body intersections
- Under the physical simulation, the design should produce a visible reduction with output speed around -15 rpm from a +30 rpm input
- The intended arrangement is a two-stage right-angle transfer with reversed output direction

Scoring points:
- correct shaft assignment and 4-gear compound structure
- right-angle layout on the expected axes and shaft centers
- stage sizes and placement consistent with a stable visible reduction
- zero intersecting solids in the authored initial state
- physical transfer under Blender rigid-body simulation
- output-stage arrangement consistent with the reversed-direction design intent
Create one solid cube.

Requirements:
- Side length exactly 20.0 mm.
- Faces aligned to global X/Y/Z axes.
- Top face lies on z=0.
- Entire cube is in z<=0.
- Center on X/Y is at (0, 0).
Create a custom nonstandard threaded bolt-and-nut pair.

Required layout:
- Create both parts as one compound.
- Bolt axis is +Z, centered at (-14, 0).
- Nut axis is +Z, centered at (+14, 0).
- Both parts must stay within z in [0, 18].

Custom thread specification:
- Nonstandard right-handed, 2-start thread
- Nominal major diameter: 7.30 mm
- Approximate root diameter: 6.00 mm
- Pitch: 1.35 mm
- Lead: 2.70 mm

End of preview. Expand in Data Studio

CAD-bench Task Payloads

This dataset contains the public task payloads for CAD-bench, an executable benchmark for language-model CAD agents. Each task directory includes:

prompt.txt: the natural-language benchmark prompt
task.toml: task metadata, difficulty, evaluator name, and expected values
gold.py: a reference Build123D solution used for validation and media generation
optional fixtures such as STEP files or Blender simulation scripts

It also includes results/cad-bench-reported-results.json, a compact reviewer-facing result artifact containing every complete 17-task row from the CAD-bench website payload used by the paper tables. The JSON separates 17 standalone model rows from 32 agent rows.

The benchmark runtime and scoring code are distributed with the accompanying anonymous submission artifact. The runtime loads this dataset by setting:

HF_TASKS_REPO_ID=CAD-bench/cad-bench-ed-2026-anonymous
HF_TASKS_REVISION=03d233433eb7ea294bd05fb00522ed98a4b77a10

Intended Use

Use this dataset with the CAD-bench runtime to evaluate CAD code-generation or agentic CAD systems. Scores are diagnostic benchmark signals; they are not certifications that generated mechanical parts are safe to manufacture or deploy.

Data Provenance

The tasks are synthetic CAD prompts and benchmark metadata authored for this benchmark. They do not contain personal data or human-subject records. The M3 socket-head task includes a STEP fixture for a standard commercial fastener, documented in the task metadata and paper license table.

Licensing

Authored benchmark code, prompts, task metadata, and reference programs are released under MIT. The M3 socket-head screw fixture is a McMaster-Carr CAD download for part 91290A111; use of that fixture is governed by McMaster-Carr's website and CAD download terms.

Limitations

The release has 17 tasks. Some simple geometry tasks are close to solved by current models, while functional assembly tasks remain difficult. The current reference implementations use Build123D, although the benchmark is intended to score submitted CAD artifacts rather than a particular modeling API.

Citation

Anonymous Author(s). CAD-bench: An Executable Benchmark for Language-Model CAD Agents. NeurIPS Evaluations & Datasets submission, 2026.

Downloads last month: 27