RoTA-Bench

This dataset is published by Qingchuan Ma. It provides both the fixed benchmark data and the public evaluation entry point.

RoTA-Bench evaluates 3D spatial reasoning through interactive projection alignment. The release contains the fixed 674-task corpus used by the benchmark, with target projections, auxiliary views, current projections, pose metadata, and sampled 2D icon points.

Two tasks from the original 676-task collection (single_00345 and single_00384) are omitted because their source icons cannot be redistributed due to copyright restrictions.

Load the dataset

from datasets import load_dataset

dataset = load_dataset("QingchuanMa/RoTA-Bench")
print(dataset)

The canonical task records are loaded as the default train split. The released test and remaining split metadata are available under splits/; use their sample_ids to select the official benchmark subsets without duplicating the 1 GB task file.

Run the evaluator

The evaluation implementation and renderer are available in this dataset repository under code/src/ and in the public GitHub repository:

https://github.com/MAC-AutoML/RoTA-Bench

git clone https://github.com/MAC-AutoML/RoTA-Bench.git
cd RoTA-Bench
git lfs pull
pip install -r requirements.txt

To run the evaluator directly from this Hugging Face repository:

git clone https://huggingface.co/datasets/QingchuanMa/RoTA-Bench
cd RoTA-Bench
pip install -r code/requirements.txt
python hf_eval.py --backend simulated --max-tasks 1

For an API-backed run, set OPENAI_API_KEY and provide --base-url and --api-model.

Hugging Face Evaluate API

For the official evaluate.load() interface, use the companion metric module:

pip install evaluate huggingface_hub

import evaluate

metric = evaluate.load("QingchuanMa/rota-bench-eval")
result = metric.compute(backend="simulated", max_tasks=1)

The module downloads this dataset and runs the released evaluator automatically. For API-backed evaluation, set OPENAI_API_KEY in the environment; the key is never stored in either repository.

Use the downloaded HF dataset as the task source in custom experiments, or use the frozen files and formal protocol documented in the GitHub README. Formal evaluation and reference search require CUDA; a one-task simulated smoke test does not require an API key.

Links

• Paper: https://arxiv.org/abs/2610.17665v1
• Hugging Face Papers: https://huggingface.co/papers/2610.17665
• Interactive demo: https://www.modelscope.cn/studios/san2333333/rotabench-anon?mode=full&app_path=%2Fportal

License and attribution

Please follow the repository license and the licenses/attribution requirements of the upstream game-icons assets. The two omitted icons are not redistributed in this release.