XiaofengAlg/SPAR-arxiv-cache

SPAR Scholarly Metadata Cache

This dataset provides a public SQLite metadata cache for SPAR, a multi-agent scholarly retrieval system. It is an operational cache, not the SPARBench evaluation dataset.

Download

mkdir -p database
hf download XiaofengAlg/SPAR-arxiv-cache arxiv_data.db.zst \
  --repo-type dataset \
  --local-dir database
zstd -d database/arxiv_data.db.zst -o database/arxiv_data.db

Direct downloads are also available from the Files and versions tab.

Contents

The repository contains one Zstandard-compressed SQLite database:

• File: arxiv_data.db.zst
• Compressed size: 402,719,736 bytes
• Uncompressed size: 2,164,948,992 bytes
• Compressed SHA-256: 7498d3ee2bf1bf84a6bfafeddc11d2e00af4ea85c6edbb7bb7419f8393458e75
• Uncompressed SHA-256: 52a7ffbb0c4f7be6a376b6e06d8dff4152dda0784fc0aae768d9d6f735e492a5
• Records: 112,581 distinct cache entries

Schema:

CREATE TABLE arxiv_docs (
    arxiv_id TEXT PRIMARY KEY,
    data TEXT NOT NULL
);

data is a JSON object. Fields vary by upstream source and may include paper identifiers, title, abstract, authors, year, venue, topic or field labels, citation metadata, and public resource URLs. The cache aggregates records collected through arXiv, PubMed, OpenAlex, Semantic Scholar, and reference expansion performed by SPAR.

Integrity and provenance

The legacy public object-store snapshot was a malformed SQLite file. This Hub version was rebuilt on 2026-08-24 by recovering intact arxiv_docs rows, preferring rows directly readable from the original table when duplicate keys had different payloads, discarding unassigned recovery fragments, and compacting the result into a new database.

Validation of this file reported:

• PRAGMA quick_check: ok
• 112,581 rows and 112,581 distinct primary keys
• zero invalid JSON payloads
• no credential-like JSON field names or known API-key markers in the audit

The cache is a best-effort snapshot rather than an authoritative or complete scholarly index. Applications should fall back to live upstream services when a record is missing and should verify time-sensitive metadata against the original source.

Licensing and responsible use

SPAR's source code is MIT-licensed, but that license does not automatically apply to third-party scholarly metadata and abstracts. This dataset therefore uses license: other. Every record remains subject to the terms and rights of its upstream source and original publisher. No license to paper full text is granted by this repository.

Do not use this snapshot as the sole basis for author identity, attribution, citation counts, medical decisions, or other high-stakes conclusions.

Citation

If this cache is used with SPAR, cite the accompanying paper:

@misc{shi2025sparscholarpaperretrieval,
  title        = {SPAR: Scholar Paper Retrieval with LLM-based Agents for Enhanced Academic Search},
  author       = {Xiaofeng Shi and Yuduo Li and Qian Kou and Longbin Yu and Jinxin Xie and Hua Zhou},
  year         = {2025},
  eprint       = {2507.15245},
  archivePrefix= {arXiv},
  primaryClass = {cs.IR},
  url          = {https://arxiv.org/abs/2507.15245}
}