license: apache-2.0

Datahub for Graph-KV

This directory contains processed datasets for retrieval-augmented generation (RAG) and Arxiv-QA tasks, used in the paper Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models.

It is organized into two main folders: rag and arxiv, and results.

📁 rag/

This folder includes preprocessed data for several commonly used RAG datasets. Each subdirectory corresponds to a different dataset split or benchmark:

• 2wiki_dev: Processed subset from the 2Wiki dataset (dev split)
• hqa_eval: HotpotQA evaluation set
• morehopqa: Morehop-QA dataset
• multihopqa: Multihop-RAG QA benchmark
• nq_eval: Narritive-QA evaluation set
• tqa_eval: TriviaQA evaluation set

All files are preprocessed and ready for use in RAG-based models.

📁 arxiv/

This folder contains the processed version of the proposed ArxivQA dataset. It is organized into the following subdirectories:

• questions/: Contains the questions used for QA tasks.
• answers/: Contains the corresponding answers.
• text_data/: Raw extracted text data from the ArXiv PDFs.
• pdfs/: Original PDF files.
  • For each QA sample, 0 folder is the central (main) paper.
  • 1 folder is the reference paper cited by the central paper.

📁 results

This folder saves the computed results of the RAG task and the Arxiv-QA task.

Reference

If you find the data useful, please consider citing our paper.

@article{wang2025graph,
  title={Graph-KV: Breaking Sequence via Injecting Structural Biases into Large Language Models},
  author={Wang, Haoyu and Wang, Peihao and Li, Mufei and Liu, Shikun and Miao, Siqi and Wang, Zhangyang and Li, Pan},
  journal={arXiv preprint arXiv:2506.07334},
  year={2025}
}