README
RealCADBench-V1.0
RealCADBench-V1.0 contains design-intent inputs and ground-truth STL shapes for evaluating AI CAD models and agents on parts and assemblies.
The test split is an evaluation collection, not a newly created train/test partition. The default all configuration combines every sample. The six other configurations provide the individual subsets without duplicating the Parquet files.
| Category | Subset | Samples |
|---|---|---|
| part | text | 568 |
| part | 2d_drawing | 236 |
| part | real_pic | 568 |
| part | real_pic_aug | 10792 |
| part | rendered_img | 373 |
| assm | assm | 167 |
| Total | 12704 |
Preview fields
| Field | Meaning |
|---|---|
| category | part or assm |
| subset | text, 2d_drawing, real_pic, real_pic_aug, rendered_img, or assm |
| input_text | Actual text input; empty string for image-only samples |
| input_images | Ordered list of input images embedded in Parquet |
| original_image | Original auxiliary image for text samples; null for other subsets |
| gt_shape | STL path relative to the repository root |
original_image is auxiliary source material for the text subset, not an additional input to the text-only evaluation task. STL geometry is distributed as original files; the table displays the STL path.
Original files and mappings
part/input/<subset>/andpart/ground_truth/<subset>/contain part data.assm/input/andassm/ground_truth/contain assembly data.part/meta_info.<subset>.jsonandassm/meta_info.jsonare the canonical mappings.- Paths in each original meta_info JSON are relative to that JSON’s directory. The preview
gt_shapeincludes thepart/orassm/prefix. - Multiple images within a sample must be kept together. Different samples may share one GT, especially in
2d_drawingandrendered_img.
The release contains only files referenced by these mappings. Transfer archives, temporary files and audit reports are not part of the dataset payload.
License
RealCADBench-V1.0 is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.
You may share and adapt the dataset, including for commercial purposes, provided that you give appropriate credit, provide a link to the license, and indicate if changes were made. See the full license terms.
Remaining augmentation assets are stored under part/input/real_pic_aug_extra/package_XXX/ and part/ground_truth/real_pic_aug_extra/package_XXX/. The metadata and Parquet paths include this layout; previously published asset paths are retained.
Contributors
500 commits
Similar repos
README
RealCADBench-V1.0
RealCADBench-V1.0 contains design-intent inputs and ground-truth STL shapes for evaluating AI CAD models and agents on parts and assemblies.
The test split is an evaluation collection, not a newly created train/test partition. The default all configuration combines every sample. The six other configurations provide the individual subsets without duplicating the Parquet files.
| Category | Subset | Samples |
|---|---|---|
| part | text | 568 |
| part | 2d_drawing | 236 |
| part | real_pic | 568 |
| part | real_pic_aug | 10792 |
| part | rendered_img | 373 |
| assm | assm | 167 |
| Total | 12704 |
Preview fields
| Field | Meaning |
|---|---|
| category | part or assm |
| subset | text, 2d_drawing, real_pic, real_pic_aug, rendered_img, or assm |
| input_text | Actual text input; empty string for image-only samples |
| input_images | Ordered list of input images embedded in Parquet |
| original_image | Original auxiliary image for text samples; null for other subsets |
| gt_shape | STL path relative to the repository root |
original_image is auxiliary source material for the text subset, not an additional input to the text-only evaluation task. STL geometry is distributed as original files; the table displays the STL path.
Original files and mappings
part/input/<subset>/andpart/ground_truth/<subset>/contain part data.assm/input/andassm/ground_truth/contain assembly data.part/meta_info.<subset>.jsonandassm/meta_info.jsonare the canonical mappings.- Paths in each original meta_info JSON are relative to that JSON’s directory. The preview
gt_shapeincludes thepart/orassm/prefix. - Multiple images within a sample must be kept together. Different samples may share one GT, especially in
2d_drawingandrendered_img.
The release contains only files referenced by these mappings. Transfer archives, temporary files and audit reports are not part of the dataset payload.
License
RealCADBench-V1.0 is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) License.
You may share and adapt the dataset, including for commercial purposes, provided that you give appropriate credit, provide a link to the license, and indicate if changes were made. See the full license terms.
Remaining augmentation assets are stored under part/input/real_pic_aug_extra/package_XXX/ and part/ground_truth/real_pic_aug_extra/package_XXX/. The metadata and Parquet paths include this layout; previously published asset paths are retained.
Similar repos
Contributors
500 commits