HongchengGao/EngiWorld-Tasks

EngiWorld Tasks

This repository contains the 1,301 tasks for EngiWorld, covering 26 engineering software platforms across six domains. It provides task definitions, evaluators, and the complete task input and reference resources, including the original engineering drawings and binary project files.

• Paper: Read the paper
• Project website: engiworld.github.io
• Code and evaluation runtime: Hongcheng-Gao/EngiWorld
• Environment images: HongchengGao/EngiWorld-Images

Browse tasks

The viewer lists all 1,301 tasks with their full instructions, paper task types, directory categories, interaction modes, software, and source file paths. main_experiment identifies the 306 tasks in the updated paper's main evaluation: 158 CLI and 148 GUI, including all 10 Open-Environment Engineering tasks.

Contents

The task/ directory follows the updated paper taxonomy and the GitHub task release. The six domains are CAD, CAE, CAM, BIM, EDA, and DCC.

Task type	Directory category	Tasks
Single-Software Execution	single-software-execution	931
Software Selection	software-selection	140
Vision-Guided Modeling	vision-guided-modeling	120
Design Optimization	design-optimization	40
Cross-Software Coordination	cross-software-coordination	60
Open-Environment Engineering	open-environment-engineering	10

Tasks include 691 CLI tasks and 610 GUI tasks. The paper's 306-task main-experiment subset is listed in task/splits/main-306.txt. The historical main-300.txt cohort is retained for reproducing earlier results.

Directory names, JSON IDs, task_file, and category use the updated names. task/taxonomy.json records the paper labels, and task/id-migration-20261009.json maps every previous ID and path to its current form. The GitHub runtime accepts both through task/aliases.json.

Download

hf download HongchengGao/EngiWorld-Tasks --repo-type dataset --local-dir EngiWorld-Tasks

This dataset includes all 2,122 task resources listed in task/assets-manifest.json, including all 120 Vision-Guided Modeling input drawings, native engineering files, simulation inputs, reference results, and the nine Altium evaluators previously stored externally. The two ANSYS wb_hertz.rst files are the complete original result files, rather than pointer-file contents. task/resources-integrity.json records sizes and SHA-256 checksums for all 2,122 completed resource files.

For evaluation, install the GitHub runtime, then use the downloaded task directory with --task-root. For example, from the runtime directory:

python -m engiworld.local_eval \
  --task-root /absolute/path/to/EngiWorld-Tasks/task \
  --task vision-guided-modeling/cli/freecad/task-01 \
  --check

For Vision-Guided Modeling, the runtime reads images from the selected task's init_file/ directory and supplies them to the model as reference-image input. The dataset download already includes these images and does not require fetching them from an older GitHub commit. From the runtime directory, download complete task data before starting evaluation:

python -m engiworld.prepare_tasks --source hf --task-root ../data/task
python -m engiworld.prepare_images --source hf --image LibreCAD2.2.0.2

The preparation command downloads this complete dataset and checks its resource hashes. Evaluation reads the local downloaded task directory and a local decompressed QCOW2 image. Prepare the Docker runner with docker pull happysixd/osworld-docker before starting a task.

Virtual machine images are provided separately in the environment image repository linked above. The task definitions, scoring rules, 1,301 task rows, and 306-task main-experiment split are unchanged by the resource completion.

License

The upstream MIT license and the existing notices in the task files are retained. Engineering software is subject to its respective licenses.

Citation

@misc{engiworld2026,
  title = {EngiWorld: What Can Frontier Agents Deliver in Professional Engineering Environments?},
  author = {Hongcheng Gao and Hailong Qu and Yu Lei and Henghui Sun and Haoyang Li and Yipeng Wei and Naihao Xue and Xiaohan Yu and Zhuo Tao and Yihe Zang and Yajiao Wang and Jingyi Tang and Yi Li and Jingjing Zhou and Jie Luo and Bohan Zeng and Chengyu Shen and Hao Jiang and Chong Chen and Bowen Qu and Olive Huang and Zeqiang Wang},
  year = {2026},
  eprint = {2609.37686},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url = {https://arxiv.org/abs/2609.37686}
}