BEIR SciDocs (orgrctera/beir_scidocs)

Overview

SciDocs is a scientific document evaluation resource originally introduced with SPECTER (Cohan et al., ACL 2020) as part of a broader SciDocs benchmark of document-level tasks (citation-related prediction, classification, and recommendation). The Benchmarking-IR (BEIR) project (Thakur et al., NeurIPS 2021 Datasets & Benchmarks) repackages the citation-prediction slice of SciDocs into the standard BEIR retrieval format: a corpus of papers, a set of queries, and qrels (query–document relevance judgments).

This Hub dataset is the CTERA-formatted release of that BEIR SciDocs test collection: each row is one retrieval query with gold relevant document IDs (and binary relevance scores) aligned to the BEIR scidocs split.

Source lineage: Allen AI SciDocs → BEIR scidocs → orgrctera/beir_scidocs.

Background

SciDocs and SPECTER

SPECTER trains a Transformer on citation graphs so that embeddings reflect inter-document relatedness—a signal that token- or sentence-level pretraining often misses. The authors introduce SciDocs, a suite of seven document-level tasks for scientific NLP. The portion used in BEIR is the citation prediction setting: given a query paper (represented by its title as the query text), retrieve papers it should cite (or that are citation-relevant in the benchmark construction)—i.e. citation-oriented retrieval over a scientific paper corpus (papers are typically drawn from Semantic Scholar–indexed literature; BEIR uses a fixed corpus and 1,000 test queries).

BEIR

BEIR aggregates 18 heterogeneous IR datasets across nine task types to study zero-shot retrieval: models are evaluated without task-specific training on each dataset. SciDocs contributes the citation prediction task type. Reported statistics for the BEIR SciDocs configuration include on the order of ~25K corpus documents, 1,000 queries, and ~4–5 relevant documents per query on average (see the BEIR paper / dataset tables).

Task

• Task type: Retrieval — specifically the BEIR SciDocs / citation-prediction setting (retrieval over a scientific paper corpus).
• Input (input): Query text (in BEIR, the title of a query paper).
• Reference output (expected_output): JSON string: list of objects {"id": "<corpus_doc_id>", "score": 1} for relevant corpus documents (binary relevance).
• Metadata: metadata.query_id is the BEIR/query identifier; metadata.split is test (BEIR SciDocs is test-only in the benchmark).

Evaluators typically embed the corpus and queries, retrieve top-k documents per query, and measure nDCG@k, MAP, Recall@k, etc., against the gold ID lists—consistent with BEIR evaluation.

Data fields

Column	Type	Description
id	string	Unique row identifier (UUID) for this Hub release.
input	string	Query string (paper title) for retrieval.
expected_output	string	JSON array of { "id": "...", "score": 1 } gold relevant documents.
metadata.query_id	string	Original query / paper id in the BEIR SciDocs pipeline.
metadata.split	string	Split name (test).

Splits: test — 1,000 query rows.

Examples

Illustrative rows from this dataset (document IDs shortened for display).

Example 1

• input: Time-Agnostic Prediction: Predicting Predictable Video Frames
• metadata.query_id: b4c80fc4b140eb08a717a446824cacb77f319166
• expected_output (structure):

[
  {"id": "0d8a5addbd17d2c7c8043d8877234675da19938a", "score": 1},
  {"id": "10b987b076fe56e08c89693cdb7207c13b870540", "score": 1},
  {"id": "385750bcf95036c808d63db0e0b14768463ff4c6", "score": 1}
]

(Additional relevant IDs appear in the full row.)

Example 2

• input: Autonomous Vehicle Navigation by Building 3 D Map and by Detecting Human Trajectory using LIDAR
• metadata.query_id: 81b14341e3e063d819d032b6ce0bc0be0917c867
• expected_output: JSON array of five { "id": "...", "score": 1 } objects (same schema as Example 1).

References

BEIR (benchmark packaging and zero-shot IR evaluation)

Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, Iryna Gurevych. BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models. NeurIPS 2021 Datasets and Benchmarks Track.

Abstract (excerpt): We introduce Benchmarking-IR (BEIR), a robust and heterogeneous evaluation benchmark for information retrieval. We leverage a careful selection of 18 publicly available datasets from diverse text retrieval tasks and domains and evaluate 10 state-of-the-art retrieval systems…

• Paper: arXiv:2104.08663
• Code / data hub: beir-cellar/beir
• Original BEIR dataset page: BeIR/scidocs on the Hub

SPECTER & SciDocs (original scientific benchmark)

Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey, Daniel Weld. SPECTER: Document-level Representation Learning using Citation-informed Transformers. ACL 2020.

Abstract (excerpt): We propose SPECTER… based on pretraining a Transformer language model on… the citation graph… we introduce SciDocs, a new evaluation benchmark consisting of seven document-level tasks ranging from citation prediction, to document classification and recommendation.

• Paper: ACL Anthology
• SciDocs data: allenai.org/data/scidocs · allenai/scidocs (GitHub)

Citation (BEIR)

@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}
}

Citation (SPECTER / SciDocs)

@inproceedings{cohan-etal-2020-specter,
  title={SPECTER: Document-level Representation Learning using Citation-informed Transformers},
  author={Arman Cohan and Sergey Feldman and Iz Beltagy and Doug Downey and Daniel Weld},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
  year={2020},
  pages={2270--2282},
  url={https://aclanthology.org/2020.acl-main.207/}
}

Notes

• This repository format (input / expected_output / metadata) is adapted for benchmark and RAG tooling; the underlying judgments follow BEIR SciDocs / SciDocs citation-prediction retrieval.
• For the raw BEIR JSONL layout (corpus / queries / qrels.tsv), see the BEIR documentation and BeIR/scidocs.