BEIR/SciDocs — Third-Party Convenience Mirror

This is not a Soroush Vahidi research dataset. It is a Parquet-format, viewer-compatible mirror of the corpus and queries splits of BeIR/scidocs, a scientific-document citation-prediction and retrieval benchmark originally built from the AllenAI SciDocs collection and redistributed by the BEIR project. It exists solely so that other repositories in this account (in particular SoroushVahidi/consistency-aware-judgments) can reference a stable, reproducible local copy of the corpus/query text. Do not cite this repository as original work — cite BEIR and/or SciDocs directly (see Citation).

Why this mirror exists

Several ranking and retrieval-consistency experiments in this account's other repositories operate on SciDocs document/query IDs and need a fixed, Dataset-Viewer-compatible snapshot to read from. Rather than vendoring a private copy, this repository re-publishes the upstream corpus and queries files as Parquet, 1:1, with no reformatting of content and no added or removed rows.

What is included

• corpus config (train split): every SciDocs document, unchanged from upstream.
• queries config (train split): every SciDocs query, unchanged from upstream.

What is not included

• Qrels (relevance judgments) are deliberately excluded. Get them from BeIR/scidocs-qrels directly — this mirror only re-packages the two file types it was needed for.
• No derived scores, rankings, or judgments of any kind. Original ranking/consistency work built on top of this data lives in SoroushVahidi/consistency-aware-judgments, not here.
• No modification to document or query text — this is a format conversion, not a curation pass.

Dataset size

Config	Split	Rows
corpus	train	25,657
queries	train	1,000

These counts were verified directly from the Parquet file footers and match the SciDocs row in BEIR's own published benchmark table (1,000 queries; ~25K corpus documents; see the full 18-dataset BEIR table in the BEIR paper).

What one row represents

• One row in corpus is one SciDocs document (a paper title + abstract/text used as a citation-prediction/retrieval candidate), identified by _id.
• One row in queries is one SciDocs query document, identified by _id, against which candidate documents are ranked (relevance judgments for this ranking are in the separate qrels repository, not here).

Dataset structure / schema

Verified against the actual Parquet columns in this mirror (not the generic multi-dataset BEIR schema description, which describes nested dict structures BEIR uses internally, not this repository's flat Parquet layout):

corpus config

Column	Type	Meaning
_id	string	Unique document identifier (matches qrels' corpus-id in the separate qrels repo)
title	string	Document title (may be empty string if not present upstream)
text	string	Document abstract/passage text
metadata	unspecified	Carried over from the upstream BEIR record; exact sub-fields not verified in this pass — inspect ds.features after loading if you depend on it

queries config

Column	Type	Meaning
_id	string	Unique query identifier (matches qrels' query-id in the separate qrels repo)
text	string	Query text
metadata	unspecified	Same caveat as above

Quickstart

from datasets import load_dataset

corpus = load_dataset("SoroushVahidi/scidocs", "corpus", split="train")
queries = load_dataset("SoroushVahidi/scidocs", "queries", split="train")

print(corpus)   # 25,657 rows
print(queries)  # 1,000 rows

To evaluate retrieval, join these against qrels loaded separately from BeIR/scidocs-qrels on _id / query-id/corpus-id.

Research use cases

Supported by the dataset's actual contents (corpus + queries only, no qrels):

• Building or testing a retrieval/embedding pipeline against a standard scientific-document corpus, pairing it with qrels pulled from the canonical BEIR qrels repository.
• Reproducing the SciDocs row of BEIR's zero-shot retrieval benchmark table.
• Supplying text inputs to consistency-aware ranking research such as SoroushVahidi/consistency-aware-judgments.

This mirror alone does not support relevance evaluation — qrels must be obtained separately.

Provenance and ownership

Original creators	SciDocs benchmark authors, redistributed via the BEIR project (Thakur, Reimers, Rücklé, Srivastava, Gurevych, NeurIPS 2021 Datasets & Benchmarks Track)
Soroush Vahidi's role	Format conversion only (JSONL → Parquet for Hugging Face Dataset Viewer compatibility). No content authored, curated, or modified.
Canonical upstream	BeIR/scidocs (corpus + queries), BeIR/scidocs-qrels (qrels, not mirrored here)
License	cc-by-sa-4.0, matching the upstream BEIR license declaration

This repository should not be cited, described, or reused as a Soroush Vahidi–authored dataset.

Relationship to related datasets

This mirror supplies raw text only. Original research that uses SciDocs identifiers for consistency-aware ranking/judgment work is published separately as SoroushVahidi/consistency-aware-judgments — that is the dataset to cite for any original contribution; this one is infrastructure.

Limitations

• No qrels: cannot be used alone for standard SciDocs retrieval-quality evaluation.
• Several upstream BEIR dataset-card template sections (curation rationale, annotation process, bias discussion, etc.) were never filled in by the original BEIR authors and are not reproduced here with invented content — see the BEIR paper and repository for authoritative details on those points.
• metadata column contents are not independently documented by this mirror; treat as opaque pass-through from upstream unless you verify otherwise.

Version history

• Current revision: hub snapshot 96caa494… (2 configs: corpus, queries). Row counts unchanged from all previously cached revisions observed locally — no evidence of upstream drift as of this audit (2026-08-18).

Related resources

• Paper: BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models (NeurIPS 2021 D&B Track)
• Code: beir-cellar/beir
• SciDocs origin: AllenAI SciDocs
• Canonical corpus/queries: BeIR/scidocs
• Canonical qrels: BeIR/scidocs-qrels
• Related original dataset: SoroushVahidi/consistency-aware-judgments

Citation

Cite the original BEIR paper — do not cite this Hugging Face repository ID:

@inproceedings{thakur2021beir,
  title={{BEIR}: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
  author={Nandan Thakur and Nils Reimers and Andreas R{\"u}ckl{\'e} and Abhishek Srivastava and Iryna Gurevych},
  booktitle={Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)},
  year={2021},
  url={https://openreview.net/forum?id=wCu6T5xFjeJ}
}

If you also want to credit the original SciDocs task/collection specifically, see the citation guidance on AllenAI's SciDocs page — not independently verified here, so no BibTeX is reproduced for it in this card.

License

cc-by-sa-4.0, matching the license declared on the upstream BeIR/scidocs repository. This license applies to the underlying SciDocs/BEIR content; it does not make this a Soroush Vahidi–licensed original work.