c1khoa/papers_retrieval_arxiv

ArXiv Papers Retrieval Dataset (2015–2025)

Overview

This dataset is designed for information retrieval (IR) research on scientific papers. It contains a large corpus of papers collected from arXiv between 2015 and 2025, along with a set of synthetic queries and relevance judgments (qrels).

The dataset can be used to evaluate retrieval systems such as:

• lexical retrieval (BM25)
• dense retrieval
• neural reranking models

Typical tasks include:

• semantic search
• document retrieval
• neural information retrieval
• academic paper recommendation

---

Dataset Structure

The dataset contains three main files:

papers_2015_2025.parquet
queries.parquet
qrels.parquet

1. Corpus

papers_2015_2025.parquet

This file contains the document corpus consisting of arXiv papers.

Field	Description
id	arXiv identifier
title	paper title
abstract	paper abstract
authors	list of authors
categories	arXiv subject categories
update_date	last update date

The corpus is filtered to include papers from the following fields:

• cs.AI
• cs.CL
• cs.IR
• cs.LG

and within the publication years 2015–2025.

---

2. Queries

queries.parquet

This file contains search queries used for retrieval evaluation.

Field	Description
query_id	unique query identifier
text	query text

Queries are automatically generated using topic templates related to areas such as:

• neural machine translation
• information retrieval
• document ranking
• semantic similarity
• neural embeddings
• question answering

Query lengths range from 1 to 6 tokens.

---

3. Relevance Judgments

qrels.parquet

This file contains query-document relevance labels.

Field	Description
query_id	query identifier
doc_id	document identifier
relevance	relevance score

Relevance scores follow a graded relevance scheme:

Score	Meaning
3	highly relevant
2	relevant
1	weakly relevant

The labels are generated by retrieving results from the arXiv search API and filtering them by:

• publication year
• subject category
• presence in the local corpus

---

Intended Use

This dataset is intended for evaluating retrieval systems including:

• BM25
• dense embedding retrieval
• approximate nearest neighbor search
• neural reranking models

Example research tasks:

• semantic search over scientific papers
• dense passage retrieval
• neural information retrieval
• academic literature search

---

Example Use

Load the dataset using Python:

import pandas as pd

corpus = pd.read_parquet("papers_2015_2025.parquet")
queries = pd.read_parquet("queries.parquet")
qrels = pd.read_parquet("qrels.parquet")

---

License

The dataset contains metadata derived from papers hosted on arXiv. Users should refer to arXiv’s terms of use for redistribution policies.

---

Citation

If you use this dataset in research, please cite the repository:

@dataset{arxiv_ir_dataset,
  title = {ArXiv Papers Retrieval Dataset},
  year = {2026},
  description = {Information retrieval dataset built from arXiv papers with queries and relevance judgments}
}

---

Acknowledgements

Paper metadata is obtained from the arXiv repository.

arXiv is maintained and operated by the Cornell University Library.