Loading dataset details…
harbor run -d au12321ua/mesh2cad-smokeharbor run -d au12321ua/mesh2cad-smokeSix self-contained mesh reconstruction smoke tasks.
harbor run -d au12321ua/mesh2cad-smokeSix self-contained mesh reconstruction smoke tasks.
harbor run -d au12321ua/mesh2cad-smokeSix self-contained Harbor tasks: reconstruct a clean solid CAD model from a noisy, incomplete triangle scan. Two reference representations across three difficulty levels.
| Difficulty | Native STEP | Official frozen polygonal reference |
|---|---|---|
| easy | case_3000584756354796 | case_7533d6cf6be1d622 |
| medium | case_3bbf75b54508489d | case_39ab405006951c2d |
| hard | case_d2d7d6d8d9bb65ec | case_58101d49791ac197 |
Use Harbor 0.23, Docker/Compose, and your chosen agent/model. Public modeling and
verification images are pinned by digest. Each task includes input, GT, binding,
evaluator, policy, licenses, and a solution/solve.sh Oracle control. No private
host reference paths or image variables are required.
The agent allowlist covers standard OpenAI, Anthropic, Gemini and OpenRouter API
hosts and PyPI. Custom endpoints need --allow-agent-host <hostname-or-IP> plus
the chosen agent's API base URL and credentials; see the repository README.
Difficulty/reference-representation labels remain in host-side task metadata,
outside the agent input. Verifier and replay stay offline.
Submit model.py --output <path.step> and model.step. The program must construct
geometry from explicit dimensions and modeling operations without mesh access,
external files, embedded precomputed STEP/BREP, or network access.
Reward gates valid solid material and bidirectional volume differences below max(1e-6 mm³, 0.001 × reference volume). Geometry-labelled topology is a soft gate. Symmetric mean point-to-surface Chamfer distance is reported in mm independently, with 50,000 area-weighted samples per side, seed 0, and no alignment.
The ordinary agent has no SimpleCADAPI SDK or skill; optional SDK/skill attachments can be compared with the same task bytes and resource settings. The common isolated replay runtime includes the fixed SDK for both experiment arms, but receives no GT or input mesh. The evaluator never executes submitted Python.
Oracle caveat: solution/ embeds known reference geometry to validate pipeline operation. It is not a compliant parametric reconstruction answer and must not be reported as model reconstruction performance.
Data derives from CADGenBench / Mecado under ODC-BY 1.0. Attribution, source hashes, and code license notices are included per task. This is a derived smoke benchmark, not the official CADGenBench score.
Six self-contained Harbor tasks: reconstruct a clean solid CAD model from a noisy, incomplete triangle scan. Two reference representations across three difficulty levels.
| Difficulty | Native STEP | Official frozen polygonal reference |
|---|---|---|
| easy | case_3000584756354796 | case_7533d6cf6be1d622 |
| medium | case_3bbf75b54508489d | case_39ab405006951c2d |
| hard | case_d2d7d6d8d9bb65ec | case_58101d49791ac197 |
Use Harbor 0.23, Docker/Compose, and your chosen agent/model. Public modeling and
verification images are pinned by digest. Each task includes input, GT, binding,
evaluator, policy, licenses, and a solution/solve.sh Oracle control. No private
host reference paths or image variables are required.
The agent allowlist covers standard OpenAI, Anthropic, Gemini and OpenRouter API
hosts and PyPI. Custom endpoints need --allow-agent-host <hostname-or-IP> plus
the chosen agent's API base URL and credentials; see the repository README.
Difficulty/reference-representation labels remain in host-side task metadata,
outside the agent input. Verifier and replay stay offline.
Submit model.py --output <path.step> and model.step. The program must construct
geometry from explicit dimensions and modeling operations without mesh access,
external files, embedded precomputed STEP/BREP, or network access.
Reward gates valid solid material and bidirectional volume differences below max(1e-6 mm³, 0.001 × reference volume). Geometry-labelled topology is a soft gate. Symmetric mean point-to-surface Chamfer distance is reported in mm independently, with 50,000 area-weighted samples per side, seed 0, and no alignment.
The ordinary agent has no SimpleCADAPI SDK or skill; optional SDK/skill attachments can be compared with the same task bytes and resource settings. The common isolated replay runtime includes the fixed SDK for both experiment arms, but receives no GT or input mesh. The evaluator never executes submitted Python.
Oracle caveat: solution/ embeds known reference geometry to validate pipeline operation. It is not a compliant parametric reconstruction answer and must not be reported as model reconstruction performance.
Data derives from CADGenBench / Mecado under ODC-BY 1.0. Attribution, source hashes, and code license notices are included per task. This is a derived smoke benchmark, not the official CADGenBench score.