Zero-To-CAD-1m

<p align="center">
<img src="assets/logo.png" alt="Zero-to-CAD" width="100%"/>
</p>

# Zero-to-CAD 1M

**One million executable, interpretable CAD construction sequences synthesized entirely without real-world data.**

<p align="center">
<img src="assets/agentic.png" alt="Zero-to-CAD agentic synthesis pipeline" width="800"/>
</p>

> **Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data**
>
> [Mohammadmehdi Ataei](https://orcid.org/0000-0002-3399-9696), [Farzaneh Askari](https://orcid.org/0000-0003-0684-1102), [Kamal Rahimi Malekshan](https://orcid.org/0009-0004-1192-4724), [Pradeep Kumar Jayaraman](https://orcid.org/0000-0001-6314-6136)
>
> Autodesk Research

## Related Resources

| Resource | Link |
|----------|------|
| 📄 **Paper** | [Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data](https://arxiv.org/abs/2604.24479) |
| 📦 **Zero-to-CAD 1M** (this dataset) | You are here |
| 📦 **Zero-to-CAD 100K** (curated subset) | [ADSKAILab/Zero-To-CAD-100k](https://huggingface.co/datasets/ADSKAILab/Zero-To-CAD-100k) |
| 🤖 **Fine-tuned Model** (Qwen3-VL-2B) | [ADSKAILab/Zero-To-CAD-Qwen3-VL-2B](https://huggingface.co/ADSKAILab/Zero-To-CAD-Qwen3-VL-2B) |
| 🗂️ **Collection** | [ADSKAILab/Zero-To-CAD](https://huggingface.co/collections/ADSKAILab/zero-to-cad) |

## Overview

Zero-to-CAD is a large-scale dataset of **1,000,000 executable CAD construction sequences** generated by an LLM operating inside a feedback-driven CAD environment. Unlike prior datasets limited to sketch-and-extrude workflows, Zero-to-CAD covers a **broad vocabulary of CAD operations** including booleans, fillets, chamfers, shells, lofts, sweeps, and patterns.

Each sample is:
- ✅ **Executable** — valid CadQuery Python code that runs and produces geometry
- ✅ **Interpretable** — named parameters, logical construction order, human-readable code
- ✅ **Geometrically validated** — multi-stage validation (topology, contiguity, complexity, export)

### How It Was Made

We frame CAD synthesis as an **agentic search problem**: an LLM is placed in a loop with a CadQuery execution environment and equipped with three tools:

1. **`execute_and_validate`** — runs code, checks geometric validity, returns structured feedback
2. **`lookup_documentation`** — TF-IDF retrieval over CadQuery API docs
3. **`grep_documentation`** — regex-based precise syntax lookup

The agent iteratively generates, executes, and repairs code across multiple turns until validation succeeds. A two-stage protocol first generates diverse part descriptions (65 categories), then synthesizes executable code for each.

<p align="center">
<img src="assets/samples.png" alt="Sample CAD models from Zero-to-CAD" width="100%"/>
</p>

## Dataset Details

### Generation Statistics

| Metric | Value |
|--------|-------|
| Total sequences | 999,633 |
| Total tokens processed | 60.2B |
| Generated tokens per design (mean / median) | 5,638 / 5,089 |
| Function calls per conversation (mean) | 4.34 |
| Zero-shot success rate | 22.3% |
| Attempts before success (mean / median) | 3.30 / 3 |
| Minimum B-Rep face count | 7 |

### Splits

| Split | Samples |
|-------|---------|
| Train | 979,633 |
| Validation | 10,000 |
| Test | 10,000 |

### Data Fields

Each sample contains:

| Field | Type | Description |
|-------|------|-------------|
| `uuid` | `string` | Unique identifier (deterministic, derived from content hash) |
| `cadquery_file` | `string` | Executable CadQuery Python source code |
| `num_faces` | `int` | Number of B-Rep faces in the final solid |
| `face_latency_ms` | `float` | Time to compute face count (ms) |
| `cadquery_ops_json` | `string` | JSON list of CAD operations used (e.g., extrude, fillet, chamfer) |
| `cadquery_ops_count` | `int` | Number of CAD operations in the construction sequence |
| `ops_latency_ms` | `float` | Time to extract operations (ms) |
| `num_renders` | `int` | Number of rendered views |
| `image_0` – `image_7` | `image` | 8 rendered views (256×256, 4 front + 4 rear) |
| `stl_file` | `bytes` | Exported STL mesh |
| `step_file` | `bytes` | Exported STEP file |

### CAD Operations Coverage

The dataset covers a broad set of operations far beyond sketch-and-extrude:

- **Sketch primitives**: rect, circle, polygon, arc, spline, slot
- **3D operations**: extrude, cut, revolve, loft, sweep
- **Modifications**: fillet, chamfer, shell, offset
- **Booleans**: union, cut, intersect
- **Patterns**: linear patterns, polar patterns, mirror
- **Features**: holes (through, blind, countersink), threads, ribs

### Part Categories

65 categories spanning structural components (brackets, gusset plates), rotational elements (pulleys, flywheels), enclosures (housings, covers), fastening hardware (clamps, spacers), and more.

## Comparison with Existing Datasets

| Dataset | Year | Size | Replayable | Readable | Operations |
|---------|------|------|------------|----------|------------|
| DeepCAD | 2021 | 178K | ✅ | ✅ | Sketch, Extrude |
| Fusion 360 Gallery | 2021 | 8.6K | ✅ | ✅ | Sketch, Extrude |
| CC3D-Ops | 2022 | 37K+ | ❌ | ❌ | Extrude/Cut, Revolve, Fillet, Chamfer (labels only) |
| CAD-Recode | 2025 | 1M | ✅ | ❌ | Sketch, Extrude |
| **Zero-to-CAD (ours)** | **2026** | **1M** | **✅** | **✅** | **Broad** (booleans, fillets, chamfers, lofts, sweeps, shells, etc.) |

## Quick Start

### Load the dataset

The train split contains ~1M rows, so use streaming to avoid downloading the entire dataset at once:

```python
from datasets import load_dataset

# Streaming mode — rows are fetched on demand
ds = load_dataset("ADSKAILab/Zero-To-CAD-1m", split="train", streaming=True)
sample = next(iter(ds))

# Get a single sample
sample = next(iter(ds))

# Display the reconstructed script
cad_code = bytes(sample["cadquery_file"]).decode("utf-8")
print(cad_code)
```

To load a smaller split entirely into memory:

```python
ds = load_dataset("ADSKAILab/Zero-To-CAD-1m", split="test")
sample = ds[0]
```

### Execute a sample

```python
import cadquery as cq

# Streaming mode — rows are fetched on demand
ds = load_dataset("ADSKAILab/Zero-To-CAD-1m", split="train", streaming=True)
sample = next(iter(ds))

# Execute the code from a sample
code = bytes(sample["cadquery_file"]).decode("utf-8")
exec(code)

# Display generated CadQuery solid
from IPython.display import display
display(result)
```

### Visualize renders

```python
for i in range(8):
img = sample[f"image_{i}"]
img.save(f"view_{i}.png")
```

## Embeddings & FAISS Index

Precomputed DINOv3 visual embeddings and a FAISS IVF-PQ index are provided for nearest-neighbor search over the full dataset. Each embedding is the mean-pooled DINOv3 feature across 8 rendered views of a CAD model.

| File | Size | Description |
|------|------|-------------|
| `embeddings/cad_gen_embeddings.npy` | 1.54 GB | DINOv3 embeddings (999,633 × 384, float32) |
| `embeddings/cad_gen_ivfpq.index` | 97.7 MB | FAISS IVF-PQ index for approximate nearest-neighbor search |
| `cad_gen_uuids.txt` | 37 MB | Row-index → item ID mapping (999,632 IDs) |
| `embeddings/cad_gen_neighbors_ivfpq.csv` | 1.73 GB | Precomputed 20 nearest neighbors for every sample |
| `embeddings/cad_gen_diverse_samples.csv` | 7.74 MB | 100K diverse subset selection metadata (cluster assignments) |

### Search for similar models

```python
import faiss
import numpy as np
from huggingface_hub import hf_hub_download

# Download index and ID mapping
index_path = hf_hub_download("ADSKAILab/Zero-To-CAD-1m", "embeddings/cad_gen_ivfpq.index", repo_type="dataset")
ids_path = hf_hub_download("ADSKAILab/Zero-To-CAD-1m", "embeddings/cad_gen_uuids.txt", repo_type="dataset")

index = faiss.read_index(index_path)
item_ids = open(ids_path).read().splitlines()

# Query with your own embedding (384-d float32)
query = np.random.rand(1, 384).astype("float32") # replace with your embedding
distances, indices = index.search(query, k=10)

for rank, (dist, idx) in enumerate(zip(distances[0], indices[0]), 1):
print(f"{rank}. {item_ids[idx]} (distance: {dist:.4f})")
```

### Load precomputed neighbors

```python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("ADSKAILab/Zero-To-CAD-1m", "embeddings/cad_gen_neighbors_ivfpq.csv", repo_type="dataset")
neighbors = pd.read_csv(path)
# Columns: item_id, neighbor_id, score, rank
```

## Example

```python
import cadquery as cq

plate_length = 60.0
plate_width = 40.0
plate_thickness = 8.0
rib_length = 20.0
rib_width = 6.0
hole_diameter = 8.0
mount_hole_dia = 3.0
fillet_radius = 2.0
chamfer_distance = 1.0
slot_width = 4.0
slot_length = 30.0

base = (
cq.Workplane('XY')
.rect(plate_length, plate_width, centered=True)
.extrude(plate_thickness)
)
base = base.edges("|Z").fillet(fillet_radius)

rib = (
cq.Workplane('XY')
.center(-plate_length/2 + rib_length/2, 0)
.rect(rib_length, rib_width, centered=True)
.extrude(plate_thickness)
)
bracket = base.union(rib)

bracket = bracket.cut(
cq.Workplane('XY').center(0, 0)
.circle(hole_diameter/2).extrude(plate_thickness + 2)
)

bracket = bracket.edges("|Z").chamfer(chamfer_distance)
result = bracket
```

## Intended Uses

- **Training CAD sequence models** — learn to generate parametric construction histories
- **Image-to-CAD reconstruction** — reconstruct editable CAD from multi-view images
- **CAD program understanding** — study the structure and semantics of construction sequences
- **Benchmarking** — evaluate generative CAD models on diverse, validated geometry

## Limitations

- All data is synthetically generated; no real-world CAD files were used
- The LLM may produce locally plausible but globally incoherent designs in some cases
- Scale conventions (max 100 units) may not match specific engineering standards
- Operation distribution reflects LLM priors, not real-world manufacturing frequency

## Citation

If you use this dataset, please cite:

```bibtex
@misc{ataei2026zerotocadagenticsynthesisinterpretable,
title={Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data},
author={Mohammadmehdi Ataei and Farzaneh Askari and Kamal Rahimi Malekshan and Pradeep Kumar Jayaraman},
year={2026},
eprint={2604.24479},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2604.24479}
}
```

## License

This dataset is released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).

<p align="center">
<img src="assets/logo.png" alt="Zero-to-CAD" width="100%"/>
</p>

# Zero-to-CAD 1M

**完全不依赖真实世界数据生成的100万条可执行、可解释的CAD构建序列。**

<p align="center">
<img src="assets/agentic.png" alt="Zero-to-CAD 智能体合成流水线" width="800"/>
</p>

> **Zero-to-CAD：无需真实数据的百万级可解释CAD程序智能体合成**
>
> [Mohammadmehdi Ataei](https://orcid.org/0000-0002-3399-9696), [Farzaneh Askari](https://orcid.org/0000-0003-0684-1102), [Kamal Rahimi Malekshan](https://orcid.org/0009-0004-1192-4724), [Pradeep Kumar Jayaraman](https://orcid.org/0000-0001-6314-6136)
>
> Autodesk Research（欧特克研究院）

## 相关资源

| 资源类型 | 链接 |
|----------|------|
| 📄 **论文** | [Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data](https://arxiv.org/abs/2604.24479) |
| 📦 **Zero-to-CAD 1M**（本数据集） | 您当前所在位置 |
| 📦 **Zero-to-CAD 100K**（精选子集） | [ADSKAILab/Zero-To-CAD-100k](https://huggingface.co/datasets/ADSKAILab/Zero-To-CAD-100k) |
| 🤖 **微调模型**（Qwen3-VL-2B） | [ADSKAILab/Zero-To-CAD-Qwen3-VL-2B](https://huggingface.co/ADSKAILab/Zero-To-CAD-Qwen3-VL-2B) |
| 🗂️ **数据集合集** | [ADSKAILab/Zero-To-CAD](https://huggingface.co/collections/ADSKAILab/zero-to-cad) |

## 数据集概览

Zero-to-CAD是一个包含**100万条可执行CAD构建序列**的大规模数据集，由在反馈驱动的CAD环境中运行的大语言模型(LLM/Large Language Model)生成。与以往仅局限于草图拉伸工作流的数据集不同，Zero-to-CAD涵盖了**丰富的CAD操作词库**，包括布尔运算、圆角、倒角、抽壳、放样、扫掠以及阵列等操作。

每条样本均具备以下特性：
- ✅ **可执行** —— 合法的CadQuery Python代码，可运行并生成几何模型
- ✅ **可解释** —— 带有命名参数、逻辑构建顺序以及人类可读的代码
- ✅ **几何验证** —— 经过多阶段验证（拓扑、连续性、复杂度、导出性）

### 数据集构建流程

我们将CAD合成建模为**智能体搜索问题**：将大语言模型置于与CadQuery执行环境的闭环中，并为其配备三类工具：

1. **`execute_and_validate`** —— 运行代码、检查几何合法性并返回结构化反馈
2. **`lookup_documentation`** —— 基于TF-IDF检索CadQuery API文档
3. **`grep_documentation`** —— 基于正则表达式的精确语法查找

该智能体经过多轮迭代，持续生成、执行并修复代码，直至验证通过。我们采用两阶段流程：首先生成多样化的零件描述（共65个类别），随后为每个描述合成可执行代码。

<p align="center">
<img src="assets/samples.png" alt="Zero-to-CAD 示例CAD模型" width="100%"/>
</p>

## 数据集详情

### 生成统计指标

| 指标 | 数值 |
|--------|-------|
| 总序列数 | 999,633 |
| 处理总Token数 | 602亿 |
| 单设计生成Token数（均值/中位数） | 5638 / 5089 |
| 单会话函数调用次数（均值） | 4.34 |
| 零样本(Zero-shot)成功率 | 22.3% |
| 成功前的尝试次数（均值/中位数） | 3.30 / 3 |
| 最小B-Rep面数 | 7 |

### 数据集划分

| 划分集 | 样本数量 |
|-------|---------|
| 训练集 | 979,633 |
| 验证集 | 10,000 |
| 测试集 | 10,000 |

### 数据字段说明

每条样本包含以下字段：

| 字段名 | 数据类型 | 描述 |
|-------|------|-------------|
| `uuid` | 字符串 | 唯一标识符（由内容哈希生成，具有确定性） |
| `cadquery_file` | 字符串 | 可执行的CadQuery Python源代码 |
| `num_faces` | 整数 | 最终实体的B-Rep面数 |
| `face_latency_ms` | 浮点数 | 计算面数的耗时（单位：毫秒） |
| `cadquery_ops_json` | 字符串 | 所用CAD操作的JSON列表（例如拉伸、圆角、倒角） |
| `cadquery_ops_count` | 整数 | 构建序列中的CAD操作数量 |
| `ops_latency_ms` | 浮点数 | 提取操作的耗时（单位：毫秒） |
| `num_renders` | 整数 | 渲染视图的数量 |
| `image_0` – `image_7` | 图像 | 8张渲染视图（分辨率256×256，4个前视图+4个后视图） |
| `stl_file` | 字节 | 导出的STL网格文件 |
| `step_file` | 字节 | 导出的STEP文件 |

### CAD操作覆盖范围

本数据集涵盖了远超草图拉伸工作流的丰富操作类型：

- **草图图元**：矩形、圆形、多边形、圆弧、样条曲线、槽
- **三维操作**：拉伸、切除、旋转、放样、扫掠
- **修改操作**：圆角、倒角、抽壳、偏移
- **布尔运算**：并集、差集、交集
- **阵列操作**：线性阵列、极坐标阵列、镜像
- **特征类型**：孔（通孔、盲孔、沉孔）、螺纹、加强筋

### 零件类别

共65个类别，涵盖结构件（支架、连接板）、回转元件（滑轮、飞轮）、外壳（壳体、盖板）、紧固件（夹具、垫片）等多种类型。

## 与现有数据集的对比

| 数据集 | 发布年份 | 规模 | 可复现 | 可读 | 操作覆盖范围 |
|---------|------|------|------------|----------|------------|
| DeepCAD | 2021 | 17.8万 | ✅ | ✅ | 草图、拉伸 |
| Fusion 360 Gallery | 2021 | 8.6K | ✅ | ✅ | 草图、拉伸 |
| CC3D-Ops | 2022 | 3.7万+ | ❌ | ❌ | 拉伸/切除、旋转、圆角、倒角（仅标签） |
| CAD-Recode | 2025 | 100万 | ✅ | ❌ | 草图、拉伸 |
| **Zero-to-CAD（本数据集）** | **2026** | **100万** | **✅** | **✅** | **丰富多样**（布尔运算、圆角、倒角、放样、扫掠、抽壳等） |

## 快速入门

### 加载数据集

训练集包含约100万条数据，建议使用流式加载以避免一次性下载全部数据集：

python
from datasets import load_dataset

# 流式加载模式 —— 按需获取数据行
ds = load_dataset("ADSKAILab/Zero-To-CAD-1m", split="train", streaming=True)
sample = next(iter(ds))

# 获取单条样本
sample = next(iter(ds))

# 展示重构的代码脚本
cad_code = bytes(sample["cadquery_file"]).decode("utf-8")
print(cad_code)

若要将较小的划分集完整加载到内存中：

python
ds = load_dataset("ADSKAILab/Zero-To-CAD-1m", split="test")
sample = ds[0]

### 执行样本代码

python
import cadquery as cq

# 流式加载模式 —— 按需获取数据行
ds = load_dataset("ADSKAILab/Zero-To-CAD-1m", split="train", streaming=True)
sample = next(iter(ds))

# 执行样本中的代码
code = bytes(sample["cadquery_file"]).decode("utf-8")
exec(code)

# 展示生成的CadQuery实体
from IPython.display import display
display(result)

### 可视化渲染视图

python
for i in range(8):
img = sample[f"image_{i}"]
img.save(f"view_{i}.png")

## 嵌入向量与FAISS索引

我们提供了预计算的DINOv3视觉嵌入向量以及FAISS IVF-PQ索引，用于在全数据集上执行最近邻搜索。每条嵌入向量为CAD模型8张渲染视图的均值池化DINOv3特征。

| 文件路径 | 大小 | 描述 |
|------|------|-------------|
| `embeddings/cad_gen_embeddings.npy` | 1.54 GB | DINOv3嵌入向量（999,633 × 384，float32格式） |
| `embeddings/cad_gen_ivfpq.index` | 97.7 MB | FAISS IVF-PQ近似最近邻搜索索引 |
| `cad_gen_uuids.txt` | 37 MB | 行索引→项目ID映射（共999,632个ID） |
| `embeddings/cad_gen_neighbors_ivfpq.csv` | 1.73 GB | 每条样本的预计算20个最近邻 |
| `embeddings/cad_gen_diverse_samples.csv` | 7.74 MB | 10万条多样化子集的选择元数据（聚类分配信息） |

### 搜索相似模型

python
import faiss
import numpy as np
from huggingface_hub import hf_hub_download

# 下载索引和ID映射文件
index_path = hf_hub_download("ADSKAILab/Zero-To-CAD-1m", "embeddings/cad_gen_ivfpq.index", repo_type="dataset")
ids_path = hf_hub_download("ADSKAILab/Zero-To-CAD-1m", "embeddings/cad_gen_uuids.txt", repo_type="dataset")

index = faiss.read_index(index_path)
item_ids = open(ids_path).read().splitlines()

# 使用自定义嵌入向量进行查询（384维float32格式）
query = np.random.rand(1, 384).astype("float32") # 替换为您的嵌入向量
distances, indices = index.search(query, k=10)

for rank, (dist, idx) in enumerate(zip(distances[0], indices[0]), 1):
print(f"{rank}. {item_ids[idx]} (距离: {dist:.4f})")

### 加载预计算的最近邻结果

python
import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("ADSKAILab/Zero-To-CAD-1m", "embeddings/cad_gen_neighbors_ivfpq.csv", repo_type="dataset")
neighbors = pd.read_csv(path)
# 列名：item_id, neighbor_id, score, rank

## 代码示例

python
import cadquery as cq

# 底板尺寸
plate_length = 60.0
plate_width = 40.0
plate_thickness = 8.0
# 加强筋尺寸
rib_length = 20.0
rib_width = 6.0
# 安装孔尺寸
hole_diameter = 8.0
mount_hole_dia = 3.0
# 圆角半径
fillet_radius = 2.0
# 倒角距离
chamfer_distance = 1.0
# 槽尺寸
slot_width = 4.0
slot_length = 30.0

base = (
cq.Workplane('XY')
.rect(plate_length, plate_width, centered=True)
.extrude(plate_thickness)
)
base = base.edges("|Z").fillet(fillet_radius)

rib = (
cq.Workplane('XY')
.center(-plate_length/2 + rib_length/2, 0)
.rect(rib_length, rib_width, centered=True)
.extrude(plate_thickness)
)
bracket = base.union(rib)

bracket = bracket.cut(
cq.Workplane('XY').center(0, 0)
.circle(hole_diameter/2).extrude(plate_thickness + 2)
)

bracket = bracket.edges("|Z").chamfer(chamfer_distance)
result = bracket

## 预期应用场景

- **训练CAD序列模型** —— 学习生成参数化构建历史
- **图像到CAD重建** —— 从多视图图像重建可编辑的CAD模型
- **CAD程序理解** —— 研究构建序列的结构与语义
- **模型基准测试** —— 在多样化且经过验证的几何模型上评估生成式CAD模型

## 局限性说明

- 所有数据均为合成生成，未使用任何真实CAD文件
- 大语言模型可能在部分场景中生成局部合理但全局不一致的设计
- 尺寸规范（最大100单位）可能不符合特定工程标准
- 操作分布反映了大语言模型的先验知识，而非真实世界的制造使用频率

## 引用格式

若您使用本数据集，请引用以下文献：

bibtex
@misc{ataei2026zerotocadagenticsynthesisinterpretable,
title={Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data},
author={Mohammadmehdi Ataei and Farzaneh Askari and Kamal Rahimi Malekshan and Pradeep Kumar Jayaraman},
year={2026},
eprint={2604.24479},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2604.24479}
}

## 开源协议

本数据集遵循[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0)协议发布。

搜集汇总

背景与挑战

背景概述

Zero-To-CAD-1m是一个包含100万个可执行、可解释的CAD构造序列的大规模数据集，完全基于合成数据生成。它通过LLM驱动的代理方法，在反馈驱动的CAD环境中生成，覆盖了广泛的CAD操作，如布尔运算、倒角和扫掠等，并经过多阶段几何验证。