---
title: omega-compositional
canonical_url: "https://www.modelscope.cn/datasets/allenai/omega-compositional"
md_url: "https://www.modelscope.cn/datasets/allenai/omega-compositional.md"
repository: allenai/omega-compositional
last_updated: 2025-06-25
license: "Apache License 2.0"
storage_size: "981 KB"
downloads: 434
stars: 0
---

# omega-compositional

> omega-compositional - allenai 在 ModelScope 开源的数据集。Compositional Math Problems

allenai/omega-compositional 是 ModelScope 魔搭社区上的数据集，存储大小 981 KB，采用 Apache License 2.0 许可。

- **Repository**: allenai/omega-compositional
- **License**: Apache License 2.0
- **Storage size**: 981 KB
- **Downloads**: 434
- **Stars**: 0
- **Last updated**: 2025-06-25

Source: https://www.modelscope.cn/datasets/allenai/omega-compositional

---

# Compositional Math Problems

This dataset combines all compositional mathematical problem settings in paper "OMEGA: Can LLMs Reason Outside the Box in Math? Evaluating Exploratory, Compositional, and Transformative Generalization" with proper train/test splits. Each compositional setting includes training data from individual mathematical domains and test data consisting of compositional problems that require cross-domain reasoning.

## Quick Start

```python
from datasets import load_dataset

# Load all compositional settings
dataset = load_dataset("allenai/omega-compositional")

# Load a specific compositional setting with train/test splits
circles_data = load_dataset("allenai/omega-compositional", "comp_circles_algebra")
train_data = circles_data["train"]  # Training problems from individual domains
test_data = circles_data["test"]    # Compositional problems requiring both domains

# Load just the training or test split
train_only = load_dataset("allenai/omega-compositional", "comp_circles_algebra", split="train")
test_only = load_dataset("allenai/omega-compositional", "comp_circles_algebra", split="test")
```

## Dataset Description

Each compositional setting combines training data from two distinct mathematical domains and provides compositional test problems that require integrating knowledge from both domains. The training problems are sourced from individual domain datasets, while test problems are specifically designed to test cross-domain reasoning capabilities.

## Citation

If you use this dataset, please cite the original work:

```bibtex
@article{sun2024omega,
  title     = {OMEGA: Can LLMs Reason Outside the Box in Math? Evaluating Exploratory, Compositional, and Transformative Generalization},
  author    = {Yiyou Sun and Shawn Hu and Georgia Zhou and Ken Zheng and Hannaneh Hajishirzi and Nouha Dziri and Dawn Song},
  journal   = {arXiv preprint arXiv:2506.18880},
  year      = {2024},
}
```

## Related Resources

- **Explorative Dataset**: See [omega-explorative](https://huggingface.co/datasets/sunyiyou/omega-explorative) for explorative reasoning challenges
- **Transformative Dataset**: See [omega-transformative](https://huggingface.co/datasets/sunyiyou/omega-transformative) for transformative reasoning challenges
- **Paper**: See the full details in [paper](https://arxiv.org/pdf/2506.18880)
- **Code Repository**: See generation code on [github](https://github.com/sunblaze-ucb/math_ood)