Qdrant/PubMed-MV

PubMed-MultiVector (PubMed-VE)

Overview

PubMed-VE is a retrieval benchmark built from PubMed abstracts using BGE-M3.

Each document is represented in all three formats produced by BGE-M3:

• ~24 million dense embeddings
• ~24 million sparse embeddings
• 8.37B multi-vector (token-level) embeddings (8,370,222,287 exactly; 350.2 tokens/document)

Statistic	Value
Documents	23,898,701
Queries	10,000
Ground truth	exact top-1000, one list per modality (dense / sparse / multi-vector)
Embeddings	BAAI/bge-m3 (dense 1024-d, sparse lexical weights, ColBERT-style multi-vector)
Corpus size	~35 TiB

Dataset Contents

The corpus is MedRAG/pubmed, which consists of 23,898,701 documents.

Each row contains the following fields:

Field	Type	Notes
id	string	MedRAG snippet ID
title	string	Article title
content	string	Abstract text
contents	string	Title + abstract
PMID	int64	PubMed ID
source_file_name	string	Source file in the MedRAG corpus
source_row_number	int64	Row within the MedRAG source file
bge_m3_dense_embedding	list<float32>	1024-D dense vector
bge_m3_sparse_embedding	struct<indices: list<uint32>, values: list<float32>>	Sparse lexical weights
bge_m3_multivector_embedding	list<list<float32>>	1024-D vector per token

Embeddings

Each PubMed abstract is represented using BGE-M3 in three retrieval modalities:

• Dense: 1024D float32 embeddings.
• Sparse: BGE-M3 lexical weights over the XLM-RoBERTa vocabulary (250,002 tokens). Indices are token ids, values are float32 weights.
• Multi-vector: A set of multi-vectors, where there is one 1024-d float32 vector per input token, ColBERT-style.

All three vectors are produced in a single forward pass per document by the same BGE-M3 checkpoint, so the dense, sparse, and multi-vector views of a document are exactly consistent with one another.

Queries and Ground Truth

The benchmark includes ground truth at k=1000 for 10,000 queries sampled from MTEB's raw bioRxiv dataset. Each query is the title of an existing bioRxiv preprint, embedded in dense, sparse, and multi-vector form, modeling a literature-review workflow in which a paper is used to retrieve related PubMed literature. Queries are short relative to the corpus, which makes the benchmark a asymmetric short-query / long-document retrieval task rather than a document-to-document similarity task. Supernova was used to compute exact, brute-force ground truth. Note that multi-vector ground truth uses MaxSim to determine relevance.

Each row in the queries files contain the following columns:

Column	Type	Description
query_id	string	Source bioRxiv DOI, e.g. 10.1101/181198
query	string	bioRxiv preprint title
doc_id	string	Source bioRxiv DOI (same value as query_id)
category	string	bioRxiv subject category
bge_m3_dense_embedding	list<float32>	1024-d, L2-normalized
bge_m3_sparse_embedding	struct<indices: list<uint32>, values: list<float32>>	Lexical weights
bge_m3_multivector_embedding	list<list<float32>>	One 1024-d vector per token
hit_ids	list<string>	Top-1000 PMIDs, best first
hit_snippet_ids	list<string>	The same 1000 hits as MedRAG snippet IDs (pubmed23n0396_19143), joining to the corpus id column
hit_scores	list<float>	Parallel similarity scores, non-increasing

Two id spaces are provided because the corpus is keyed by the MedRAG snippet ID (id) while PubMed itself is keyed by PMID.

Intended Uses

PubMed-VE is intended for evaluating and comparing dense, sparse, multi-vector, hybrid, and multi-stage retrieval systems over a common scientific corpus.

Because the three modalities share a corpus, a query set, and an embedding model, differences in measured quality are attributable to the retrieval method rather than to the data or the encoder. Concretely, it supports:

• Recall@k against exact ground truth for an ANN index, per modality, at any k up to 1000.
• Modality comparison -- dense vs. sparse vs. multi-vector on identical inputs.
• Hybrid retrieval -- score fusion, reciprocal-rank or other result fusion, and multi-stage pipelines (e.g. dense or sparse candidate generation followed by MaxSim reranking), each with an exact reference to measure against.

Licensing

We release this dataset under the Apache 2.0 license.

Acknowledgments

We thank HuggingFace for providing a grant to offset a portion of the cost of hosting this dataset. Additionally, this benchmark would not exist without the upstream work it builds on:

• Xiong, Jin, Lu, and Zhang for the MedRAG PubMed corpus, and the U.S. National Library of Medicine for maintaining and distributing PubMed.
• bioRxiv and the preprint authors whose titles form the query set, and the MTEB team for the raw_biorxiv collection.
• BAAI for releasing BGE-M3.