CADFS Dataset

CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models

🚀 Project Page 💻 Code 🤗 Model

A large-scale dataset for parametric CAD model generation from text descriptions and multi-view images. Models are represented as FeatureScript programs, enabling direct import into Onshape environment.

This dataset was used to train and evaluate CADFS, a fine-tuned Qwen2-VL-2B multimodal language model for text-to-CAD and image-to-CAD generation.

---

Data Description

dataset/ — Processed Data

File	Description
featurescript_fp.zip	Full-precision processed FeatureScript programs.
featurescript_rp.zip	Used in training and evaluation. Reduced-precision variant with floating-point values rounded to 2 decimal places. Produces a more compact token representation suitable for language model training. Note that rounding may break compilability for some models, while others are only compilable in reduced precision.
text_annotations.zip	Natural language annotations describing the geometry, topology, and design intent of each model, obtained with GPT-OSS-120b.
step.zip	STEP files with B-rep geometry. Rendered from featurescript_fp.zip where possible, or featurescript_rp.zip otherwise.

raw/ — Raw Source Data

Source data used to create the processed FeatureScript representation.

File	Description
featurescript_raw.zip	Unprocessed FeatureScript programs.
sketch_raw.zip	Sketch metadata extracted from FeatureScript feature trees.
step_abc.zip	STEP files extracted from the ABC dataset, used in both training and evaluation.
multiview_images_abc.zip	Multi-view images rendered from step_abc.zip, used in both training and evaluation.

test_data/ — Evaluation Benchmarks

Minimal data required to perform evaluation, without the need to download the full dataset. A .json metadata file and a .zip archive are provided for each benchmark:

Benchmark	Contents
CADFS	jsonl, fs, annotations, images, step
DeepCAD	jsonl, fs, annotations, images, step
CADParser	jsonl, images, step

train_data/ — Training Splits

Training data is split into two stages following the CADFS two-stage fine-tuning strategy, and two input modalities each. We filtered out duplicates and kept only samples whose input and output fit within a context size of 8192 tokens.

File	Stage	Modality	Description
stage1_txt_train.jsonl	Stage 1	Text	Pre-training on text-to-FeatureScript generation
stage1_img_train.jsonl	Stage 1	Image	Pre-training on image-to-FeatureScript generation
stage2_txt_train.jsonl	Stage 2	Text	Fine-tuning with high-quality curated text pairs
stage2_img_train.jsonl	Stage 2	Image	Fine-tuning with high-quality curated image-program pairs

For training and evaluation we use .jsonl data format. Each .jsonl line follows the format:

Image input:

{
  "messages": [
    {
      "role": "system",
      "content": "You are CAD code generation model."
    },
    {
      "role": "user",
      "content": "<image>Generate a CAD model using FeatureScript framework..."
    },
    {
      "role": "assistant",
      "content": "FeatureScript 1511;\n..."
    }
  ],
  "images": ["path/to/0085/00858269.png"],
  "cad_file_id": "00858269"
}

Text input:

{
  "messages": [
    {
      "role": "system",
      "content": "You are CAD code generation model."
    },
    {
      "role": "user",
      "content": "Step 1 - Sketch\nCreate a new sketch on the default top plane..."
    },
    {
      "role": "assistant",
      "content": "FeatureScript 1511;\n..."
    }
  ],
  "cad_file_id": "00858269"
}

---

Usage

For usage examples, inference code, and FeatureScript processing pipeline, see the CADFS GitHub repository.

License

This dataset is released under CC BY 4.0.
It may be used for any purpose, including commercial, with attribution.

---

Citation

If you use this dataset in your research, please cite:

@inproceedings{pyatov2026cadfs,
    title      = {{{CADFS}}: A Big {{CAD}} Program Dataset and Framework for Computer-Aided Design with Large Language Models},
    shorttitle = {{{CADFS}}},
    booktitle  = {2026 {{IEEE}}/{{CVF Conference}} on {{Computer Vision}} and {{Pattern Recognition}} ({{CVPR}})},
    author     = {Vladislav Pyatov and Gleb Bobrovskikh and Saveliy Galochkin and Nikita Boldyrev and Oleg Voynov and Alexander Filippov and Gonzalo Ferrer and Peter Wonka and Evgeny Burnaev},
    year       = 2026,
    month      = jun,
    langid     = {english}
}