Graph-COM/GraphKV

---
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](http://www.arxiv.org/pdf/2506.07334).

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](https://arxiv.org/abs/2011.01060) dataset (dev split)
- `hqa_eval`: [HotpotQA](https://arxiv.org/abs/1809.09600) evaluation set
- `morehopqa`: [Morehop-QA](https://arxiv.org/abs/2406.13397) dataset
- `multihopqa`: [Multihop-RAG](https://arxiv.org/abs/2401.15391) QA benchmark
- `nq_eval`: [Narritive-QA](https://arxiv.org/abs/1712.07040) evaluation set
- `tqa_eval`: [TriviaQA](https://arxiv.org/abs/1705.03551) 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.

---