FutureCAD

Modeling histories and text descriptions for Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding (ICML 2026).

Paper · Code · BRepGround

Contents

Each row is one CAD model with its original feature definitions and ordered construction sequence, paired with an overall and a detailed text description. Histories cover sketches and solid operations, including extrude, revolve, fillet, chamfer and shell.

Split	Models
train	129,020
validation	7,167
test	7,168

Splits are disjoint by model ID. splits.json uses train, val, and test; Hugging Face Datasets names the validation split validation.

Load with Hugging Face Datasets

pip install -r requirements.txt

import json
from datasets import load_dataset

dataset = load_dataset("jhlee11/FutureCAD")
row = dataset["train"][0]
history = json.loads(row["history"])
overall_description = row["overall_description"]
detail_description = row["detail_description"]
for operation in history["sequence"]:
    feature = history["features"][operation["feature_id"]]
    print(operation["type"], feature["name"])

Use load_dataset("jhlee11/FutureCAD", split="train", streaming=True) to iterate without downloading the entire split first.

Fields

Column	Meaning
id	Eight-character CAD model ID; retain leading zeros
is_advanced	Membership in the released FutureCAD Advanced subset
num_operations	Number of construction-sequence entries, including sketches
operation_types	Ordered operation types; aligned with sequence
history	UTF-8 JSON string with features and sequence
overall_description	Abstract description of the overall shape and structure
detail_description	Detailed description of geometry and scale

Parse history with json.loads. features maps feature IDs to their definitions; sequence lists operations in construction order, with each feature_id referencing that map. Use is_advanced to select the Advanced subset.

Both description levels are used for supervised fine-tuning. The main text-to-CAD evaluation uses detail_description.

Export individual JSON files

Download export_data.py and requirements.txt from this repository, then run:

python export_data.py --split all --output datasets/futurecad_history

For one model, add --id <eight-character-id>. For an already downloaded repository, add --source /path/to/FutureCAD.

To export the test descriptions as an ID-to-text JSON for text-to-CAD evaluation:

python export_data.py --split test --format descriptions --output datasets/futurecad/test

This writes datasets/futurecad/test/descriptions.json using detail_description. Use --description-level overall to export the abstract descriptions instead.

To render a model, convert its modeling-history JSON into a CadQuery program, then execute the program and render the resulting CAD geometry.

data/train-*.parquet
data/validation-*.parquet
data/test-*.parquet
splits.json
export_data.py
requirements.txt
README.md
LICENSE.md

License

The dataset is released under CC BY 4.0.

Citation

@misc{li2026highfidelitycadgenerationllmdriven,
  title={Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding},
  author={Jiahao Li and Qingwang Zhang and Qiuyu Chen and Guozhan Qiu and Yunzhong Lou and Xiangdong Zhou},
  year={2026},
  eprint={2603.11831},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.11831}
}