TREC-COVID

Overview

TREC-COVID is a biomedical retrieval benchmark built during the COVID-19 pandemic over the CORD-19 corpus of coronavirus research papers. It is one of the 18 zero-shot datasets in BEIR and, within this vault, the dataset that appears most often when a retrieval method is being demonstrated on a specialized domain.

Its distinguishing feature among BEIR subsets is judgment depth. Where most subsets are shallow and binary, TREC-COVID carries graded relevance and hundreds of judgments per topic — the product of an organized TREC assessment effort rather than incidental labels.

TREC-COVIDNatural Questions
Queries504,352
Judgments per query~493.5~1.2
RelevanceGradedBinary

Corpus size is reported as 171K documents in Improving Search Ranking with Few-Shot Prompting of LLMs.

Why the Judgment Depth Matters

Deep graded judgments make TREC-COVID one of the few BEIR subsets where nDCG@10 measures roughly what it claims to. Elsewhere in the suite, unjudged-but-relevant documents are common and recall-oriented conclusions understate real quality — the caveat recorded in BEIR and MS MARCO.

The cost is the other axis: 50 queries. Per-dataset differences of a point or two on TREC-COVID are not obviously meaningful, and averaging it into a BEIR mean silently gives one deeply judged 50-query set the same weight as a shallow 4,352-query one. See Improving Zero-Shot Ranking with Vespa Hybrid Search, which makes this asymmetry its argument for reading per-dataset tables rather than averages.

Reported Results in This Vault

TREC-COVID as a progression across the Vespa zero-shot series, nDCG@10:

MethodScoreSource
Published BEIR BM250.656part two
Tuned BM25 (k1=0.9, b=0.4, title+text)0.690part two
Distilled 22M ColBERT rerank0.658part two
Hybrid BM25 + ColBERT0.750part two
PROMPTAGATOR (137B FLAN, few-shot)0.762part two
22M cross-encoder on synthetic data80.2few-shot prompting post

The third article quotes its baselines as 70.0 (BM25) and 76.0 (hybrid) where part two’s table gives 0.690 and 0.750. The figures differ slightly between the two posts; both are recorded as stated.

Elsewhere: Announcing the Vespa ColBERT Embedder reports 0.8003 compressed vs 0.7939 uncompressed on trec-covid, and ColBERT-Zero - To Pre-train Or Not To Pre-train ColBERT Models and What is ColBERT and Late Interaction and Why They Matter in Search both include it in their BEIR tables.

CORD-19 as a Deployment

The corpus is not only a benchmark — it is indexed and searchable at cord19.vespa.ai with selectable ranking strategies, which makes it one of the few places in the vault where a benchmark’s ranking profiles can be compared interactively rather than only read off a table. See Vespa - Ranking Without Labels on CORD-19.

  • BEIR — the suite it belongs to
  • MS MARCO — the shallow-judgment contrast
  • Natural Questions — the other end of the queries/judgments tradeoff

Articles

Source