BEIR TREC-COVID (orgrctera/beir_trec_covid)

Overview

TREC-COVID is a biomedical ad hoc retrieval benchmark organized by NIST as part of the TREC program during the COVID-19 pandemic. Participants searched a large corpus of scientific papers about COVID-19 and related viruses; assessors labeled which articles were relevant to each topic (information need). The document collection is derived from CORD-19 (COVID-19 Open Research Dataset), maintained by the Allen Institute for AI and partners.

BEIR (Benchmarking-IR) repackages TREC-COVID as one of its Bio-Medical IR tasks for zero-shot evaluation of retrieval models: same corpus, queries, and pooled relevance judgments, in a uniform JSONL/qrels format.

This Hub release follows the same tabular layout as other CTERA BEIR exports: one row per query, with input as the query string and expected_output as a JSON string of { "id", "score" } pairs (qrels-style graded relevance: TREC-COVID uses levels 1 and 2).

Task

• Task type: Retrieval (passage- or document-level ranking over the CORD-19-derived corpus) in the BEIR evaluation setting.
• Input: A natural-language topic (information need) in input.
• Supervision: expected_output lists relevant corpus document IDs with relevance scores (1 or 2), aligned with the BEIR qrels file for TREC-COVID.

Systems are typically scored with standard IR metrics (e.g. nDCG@k, MAP, Recall@k) after retrieving from the fixed corpus. The official TREC challenge also considered residual-collection evaluation in later rounds; BEIR uses the released test queries and judgments for reproducible benchmarking.

Background

TREC-COVID

The shared task was created to improve search over rapidly growing COVID-19 scientific literature for researchers, clinicians, and policymakers. It ran in multiple rounds (2020), with topics and judgments evolving over time. The BEIR snapshot uses the standard packaged split: 50 topics and the CORD-19-based corpus at the version used in the BEIR preprocessing (on the order of ~171K documents in the BEIR distribution; see the BEIR datasets table).

CORD-19

Corpus articles come from CORD-19: full-text and metadata from PubMed Central, bioRxiv, medRxiv, and other sources, aggregated into a research dataset for COVID-19 and coronavirus-related work.

BEIR

Thakur et al. (2021) introduced BEIR: a heterogeneous benchmark for zero-shot evaluation of retrieval models across 18 public datasets. TREC-COVID is one of the biomedical tasks, alongside e.g. NFCorpus and BioASQ.

• Code / resources: UKPLab/beir · BeIR/trec-covid on Hugging Face

Data fields

Column	Type	Description
id	string	Unique row identifier (UUID).
input	string	The query / topic text (information need).
expected_output	string	JSON array of { "id": "<doc_id>", "score": <1|2> } for relevance judgments (BEIR qrels).
metadata.query_id	string	Source query / topic id in the BEIR / TREC pipeline.
metadata.split	string	test (BEIR evaluation split for this dataset).

Splits

Split	Items
test	50

Examples

Illustrative rows from this dataset (IDs and text as stored). Relevance lists are large; Example 1 shows a short prefix of expected_output—the stored value contains the full list.

Example 1 — metadata.query_id 10

• input: has social distancing had an impact on slowing the spread of COVID-19?
• expected_output (prefix; truncated for display):

[{"id": "01f5mvsc", "score": 2}, {"id": "07yfqbvp", "score": 1}, {"id": "0a49okho", "score": 2}, {"id": "0b6dsdct", "score": 1}, {"id": "0dd6alkp", "score": 2}

• metadata.split: test

Example 2 — metadata.query_id 19

• input: what type of hand sanitizer is needed to destroy Covid-19?
• expected_output (prefix; truncated for display):

[{"id": "wzxqrfhu", "score": 1}, {"id": "02azobp3", "score": 2}, {"id": "0guonj3j", "score": 1}, {"id": "0macgbcn", "score": 2}, {"id": "19hy045v", "score": 1}

• metadata.split: test

References

TREC-COVID (shared task)

Abstract (JAMIA): The authors describe the rationale and structure of the TREC-COVID information retrieval shared task for COVID-19, including topic development, assessment, and evaluation considerations for pandemic-era biomedical search.

• TREC-COVID: Rationale and Structure of an Information Retrieval Shared Task for COVID-19 — JAMIA (2020) · PMC full text
• TREC-COVID hub — NIST · Data description

Follow-up overview

• Searching for Answers in a Pandemic: An Overview of TREC-COVID — NIST publication index

CORD-19 (corpus)

• CORD-19: The COVID-19 Open Research Dataset — Allen AI
• Wang et al., CORD-19: The COVID-19 Open Research Dataset — arXiv:2004.10706 cs.DL

BEIR (benchmark including TREC-COVID)

Abstract (arXiv:2104.08663): 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…

• BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models — arXiv:2104.08663

Citation

If you use TREC-COVID, cite the task overview (and the corpus as appropriate). If you use the BEIR packaging, cite BEIR:

@article{voorhees2020trec,
  title={TREC-COVID: rationale and structure of an information retrieval shared task for COVID-19},
  author={Voorhees, Ellen M and Soboroff, Ian and Alam, Tasmeer and Roberts, Kirk and Hersh, William and Demner-Fushman, Dina and Bedrick, Steven and Lo, Kyle and Wang, Lucy Lu},
  journal={Journal of the American Medical Informatics Association},
  volume={27},
  number={9},
  pages={1431--1437},
  year={2020},
  publisher={Oxford University Press},
  doi={10.1093/jamia/ocaa091}
}

@article{thakur2021beir,
  title={BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models},
  author={Thakur, Nandan and Reimers, Nils and R{\"u}ckl{\'e}, Andreas and Srivastava, Abhishek and Gurevych, Iryna},
  journal={arXiv preprint arXiv:2104.08663},
  year={2021}
}

License

This dataset card describes a BEIR-formatted export of TREC-COVID / CORD-19-derived content. Respect the original licenses of CORD-19 articles and NIST/TREC redistribution terms when using the data. The YAML license: mit here refers to this card’s packaging metadata where applicable; verify upstream terms for your use case.