“Engineering dataset” is too broad to name a useful training set. A model that classifies mechanical parts needs different examples from one that reads dimensions on a drawing or edits a CAD solid. Start with the task, then choose the source files, labels, and relationships that make the task measurable.
For mechanical CAD, this usually means deciding whether the model sees geometry, drawings, design steps, or a combination of them. It also means checking what the dataset treats as ground truth.
Match the representation to the task
| Task | Inputs and targets to look for | Example source |
|---|---|---|
| Geometric learning | CAD surfaces or solids with the geometric target used in evaluation | ABC provides parametric surfaces and benchmarks for geometric learning. |
| CAD program generation | A target shape linked to the sequence or program that made it | DeepCAD provides construction sequences. Zero-to-CAD 100K provides synthetic CadQuery programs with rendered views and STEP files. |
| Drawing-to-solid work | Drawings paired with reference solids, plus enough structure to compare the result | OrlaMechCAD-Nano links DWG sheets, STEP solids, dimensions, and views for 14 components. |
| CAD generation evaluation | Inputs and held-out answers for a defined task | CADGenBench publishes generation and editing inputs while keeping its reference answers private. |
The last row is an important boundary. CADGenBench describes its public files as benchmark inputs, not a training corpus. Downloading a benchmark does not mean you have the target outputs needed to train the same task.
File count is a poor substitute for coverage
Before comparing dataset sizes, check the variation that matters to your task. A hundred similar brackets may be less useful than a smaller collection that includes shafts, housings, assemblies, and different drawing styles. That is a task-specific judgment, not a general ranking of datasets.
Look at the source of each example and the relationships between examples. If two rendered views or revised files come from the same part, keep that relationship when you make train and evaluation splits. Otherwise, the model may see nearly the same geometry on both sides of the split. Also check units, coordinate systems, invalid solids, missing links, and whether a label was measured, inferred, or generated.
The Hugging Face dataset card guidance calls for documenting contents, intended uses, limitations, and license metadata. A good card makes these checks easier, but you still need to inspect records and files for your specific task.
A connected example from Orla
In OrlaMechCAD-Nano, the drawing and solid are related through explicit catalog records. The dataset includes DWG and DXF drawings, STEP solids, PDF sheets, dimensions, face-level geometry, and camera metadata for rendered views. That structure lets a team prototype tasks that cross 2D and 3D representations without guessing which files belong together.
Nano is a 14-component sample. It supports format evaluation and qualitative experiments. It does not prove training performance or cover every mechanical design domain. Orla can discuss a larger collection based on the part families, native formats, context, and usage rights your project needs.
If you are still choosing a public source, compare the CAD model datasets and their limits. If you have a defined task and need project-level data, share the brief with Orla.