HyperAlign-Bench

This repository contains the released HyperAlign-Bench data bundle for the paper Hypergraph as Language.

HyperAlign-Bench is a benchmark for evaluating high-order association modeling in hypergraph-language alignment. It provides vertex classification and hyperedge classification tasks under the same question-answering protocol used by Hyper-Align.

The released bundle includes processed hypergraph tensors, task samples, prebaked samples, Qwen3-Embedding-0.6B features, and overview features for five datasets.

Files

arxiv_hg/
cora_cc/
pubmed/
dblp/
imdb/

Each dataset directory follows this structure:

processed_data.pt
meta.json
samples/
embeddings/qwen3emb_0.6b/
overview/qwen3emb_0.6b/

• processed_data.pt contains the processed hypergraph data.
• meta.json contains dataset metadata.
• samples/ contains the VC and HEC task samples.
• embeddings/qwen3emb_0.6b/ contains node and hyperedge text features.
• overview/qwen3emb_0.6b/ contains overview features used by the released Hyper-Align pipeline.

For code, setup, and evaluation instructions, see the GitHub repository:

https://github.com/Mengqi-Lei/Hypergraph-as-Language

License

The HyperAlign-Bench release is distributed under the Apache License 2.0.

Qwen3-Embedding-0.6B is not redistributed here as a model. Users must comply with the license and terms of the upstream model repository:

• Qwen/Qwen3-Embedding-0.6B

Citation

If you use this dataset, please cite:

@article{lei2026hypergraph,
  title={Hypergraph as Language},
  author={Lei, Mengqi and Xie, Guohuan and Ying, Shihui and Du, Shaoyi and Yong, Jun-Hai and Li, Siqi and Gao, Yue},
  journal={arXiv preprint arXiv:2605.21858},
  year={2026}
}

Links

• Paper: https://arxiv.org/abs/2605.21858
• Code: https://github.com/Mengqi-Lei/Hypergraph-as-Language