Add dataset card (inputs)
Browse files
README.md
ADDED
|
@@ -0,0 +1,143 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: odc-by
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
pretty_name: CADGenBench (Inputs)
|
| 6 |
+
task_categories:
|
| 7 |
+
- image-to-3d
|
| 8 |
+
- text-to-3d
|
| 9 |
+
tags:
|
| 10 |
+
- cad
|
| 11 |
+
- 3d
|
| 12 |
+
- step
|
| 13 |
+
- mechanical-engineering
|
| 14 |
+
- benchmark
|
| 15 |
+
- cad-generation
|
| 16 |
+
- cad-editing
|
| 17 |
+
size_categories:
|
| 18 |
+
- n<1K
|
| 19 |
+
annotations_creators:
|
| 20 |
+
- expert-generated
|
| 21 |
+
language_creators:
|
| 22 |
+
- expert-generated
|
| 23 |
+
multilinguality:
|
| 24 |
+
- monolingual
|
| 25 |
+
source_datasets:
|
| 26 |
+
- original
|
| 27 |
+
---
|
| 28 |
+
|
| 29 |
+
# CADGenBench (Inputs)
|
| 30 |
+
|
| 31 |
+
Public inputs for the **CADGenBench** benchmark, which measures how well AI
|
| 32 |
+
systems produce correct 3D mechanical parts as STEP files. This repository
|
| 33 |
+
holds the task inputs only; the ground truth is withheld in a separate private
|
| 34 |
+
repository so that the leaderboard's evaluation is the single source of truth.
|
| 35 |
+
|
| 36 |
+
- **Leaderboard Space**: [`HuggingAI4Engineering/cadgenbench-leaderboard`](https://huggingface.co/spaces/HuggingAI4Engineering/cadgenbench-leaderboard)
|
| 37 |
+
- **Browse the tasks**: the [Tasks tab](https://huggingface.co/spaces/HuggingAI4Engineering/cadgenbench-leaderboard) on the Space (thumbnails, search, generation/editing filter, per-task detail)
|
| 38 |
+
- **Benchmark code**: [`github.com/huggingface/cadgenbench`](https://github.com/huggingface/cadgenbench)
|
| 39 |
+
- **Submissions + results**: [`HuggingAI4Engineering/cadgenbench-submissions`](https://huggingface.co/datasets/HuggingAI4Engineering/cadgenbench-submissions)
|
| 40 |
+
- **Ground truth**: [`HuggingAI4Engineering/cadgenbench-data-gt`](https://huggingface.co/datasets/HuggingAI4Engineering/cadgenbench-data-gt) (private)
|
| 41 |
+
|
| 42 |
+
## Dataset Summary
|
| 43 |
+
|
| 44 |
+
CADGenBench contains **81 fixtures** of mechanical parts split across two
|
| 45 |
+
tasks:
|
| 46 |
+
|
| 47 |
+
- **Generation (49)**: reproduce the part as an accurate 3D solid from an
|
| 48 |
+
engineering drawing.
|
| 49 |
+
- **Editing (32)**: apply a described change to an existing STEP solid.
|
| 50 |
+
|
| 51 |
+
Each submission is one `output.step` per fixture. Outputs are scored against
|
| 52 |
+
the private ground truth by the CAD Score pipeline: a hard validity gate
|
| 53 |
+
followed by a weighted mean of three orthogonal metrics (shape similarity,
|
| 54 |
+
interface match, topology match). See the
|
| 55 |
+
[metric definitions](https://github.com/huggingface/cadgenbench/blob/main/docs/metrics.md).
|
| 56 |
+
|
| 57 |
+
## Supported Tasks and Leaderboards
|
| 58 |
+
|
| 59 |
+
The benchmark is tool-agnostic: a submission is one STEP file per fixture,
|
| 60 |
+
produced by any system (one LLM, several, a script, or by hand). Submit and
|
| 61 |
+
view results on the
|
| 62 |
+
[leaderboard Space](https://huggingface.co/spaces/HuggingAI4Engineering/cadgenbench-leaderboard);
|
| 63 |
+
the full submission contract (zip layout, `meta.json`, validity gate) is in
|
| 64 |
+
[`docs/benchmark/submission.md`](https://github.com/huggingface/cadgenbench/blob/main/docs/benchmark/submission.md).
|
| 65 |
+
|
| 66 |
+
## Languages
|
| 67 |
+
|
| 68 |
+
Task prompts and descriptions are in English.
|
| 69 |
+
|
| 70 |
+
## Dataset Structure
|
| 71 |
+
|
| 72 |
+
### Data Instances
|
| 73 |
+
|
| 74 |
+
One directory per fixture, named by its numeric id. There are two shapes:
|
| 75 |
+
|
| 76 |
+
```
|
| 77 |
+
# Generation fixture
|
| 78 |
+
<id>/
|
| 79 |
+
├── description.yaml # prompt + metadata
|
| 80 |
+
└── input.png # the engineering drawing (input2.png, ... when multi-image)
|
| 81 |
+
|
| 82 |
+
# Editing fixture
|
| 83 |
+
<id>/
|
| 84 |
+
├── description.yaml # prompt + metadata
|
| 85 |
+
├── edit_description.txt # the requested change, as an instruction
|
| 86 |
+
├── input.step # the starting solid to edit
|
| 87 |
+
├── input.mesh.npz # trusted watertight mesh sidecar for input.step
|
| 88 |
+
└── renders/ # iso / front / top / right PNGs of the starting solid
|
| 89 |
+
```
|
| 90 |
+
|
| 91 |
+
### Data Fields
|
| 92 |
+
|
| 93 |
+
`description.yaml` carries:
|
| 94 |
+
|
| 95 |
+
| Field | Type | Description |
|
| 96 |
+
|---|---|---|
|
| 97 |
+
| `description` | string | The task prompt. |
|
| 98 |
+
| `task_type` | `"generation"` \| `"editing"` | The task family. |
|
| 99 |
+
| `input_files` | list of strings | The input files the task declares (e.g. `input.png`, `input2.png`, or `input.step`). |
|
| 100 |
+
| `input_type` | `"text+image"` \| `"text+step"` | Modality of the inputs. |
|
| 101 |
+
|
| 102 |
+
### Data Splits
|
| 103 |
+
|
| 104 |
+
There is no train/test split — all 81 fixtures are an evaluation set
|
| 105 |
+
(49 generation, 32 editing). The fixture set may grow over time; consumers
|
| 106 |
+
should discover fixtures at runtime rather than hardcoding the list.
|
| 107 |
+
|
| 108 |
+
## Dataset Creation
|
| 109 |
+
|
| 110 |
+
### Source Data
|
| 111 |
+
|
| 112 |
+
The underlying CAD geometry is sourced from [Mecado](https://www.mecado.com).
|
| 113 |
+
Fixtures are real mechanical parts with mating interfaces (locating jigs, bolt
|
| 114 |
+
patterns, slots). For each part, an engineering drawing (generation) or a
|
| 115 |
+
starting solid plus an edit instruction (editing) forms the public input; the
|
| 116 |
+
solved solid and its interface sub-volumes are held privately as ground truth.
|
| 117 |
+
|
| 118 |
+
### Curation Rationale
|
| 119 |
+
|
| 120 |
+
The inputs are released publicly so contestants see exactly what they are
|
| 121 |
+
solving, while the ground truth stays private so the leaderboard's server-side
|
| 122 |
+
evaluation is the only path to a score.
|
| 123 |
+
|
| 124 |
+
## Considerations for Using the Data
|
| 125 |
+
|
| 126 |
+
This is a benchmark **input** set, not a training corpus. Ground truth is not
|
| 127 |
+
included here. To participate, generate one `output.step` per fixture and
|
| 128 |
+
submit through the leaderboard Space.
|
| 129 |
+
|
| 130 |
+
## Additional Information
|
| 131 |
+
|
| 132 |
+
### Dataset Curators
|
| 133 |
+
|
| 134 |
+
The CADGenBench team ([HuggingAI4Engineering](https://huggingface.co/HuggingAI4Engineering)).
|
| 135 |
+
|
| 136 |
+
### Licensing Information
|
| 137 |
+
|
| 138 |
+
Released under the [Open Data Commons Attribution License (ODC-BY)](https://opendatacommons.org/licenses/by/1-0/).
|
| 139 |
+
|
| 140 |
+
### Acknowledgements
|
| 141 |
+
|
| 142 |
+
CAD geometry sourced from [Mecado](https://www.mecado.com). Thanks to the
|
| 143 |
+
Mecado team.
|