DictyCite

The archival and citable release is available on Zenodo: 10.5281/zenodo.20308282. The data files hosted here are identical to those in the corresponding Zenodo release.

For citation, please cite the paper together with the Zenodo DOI. Future dataset releases will continue to be archived on Zenodo and mirrored here for convenient access.

DictyCite is a benchmark for domain-specific biological literature retrieval, built from dictyBase curator notes. Each query is a compact, curator-written claim from a Dictyostelium discoideum gene summary page, and the relevance judgments are the PubMed articles the curator cited for it.

• Paper: Yun Wang, Gad Shaulsky, Tomaž Curk, Blaž Zupan. Benchmarking Literature Retrieval for a Model Organism: A Dictyostelium Case Study. Discovery Science (DS 2026), Springer, 2026 (to appear). Preprint: arXiv:2610.03130
• Canonical dataset: 10.5281/zenodo.20308282
• Code and reproduction: https://github.com/fulaibaowang/dictycite

At a glance: 1,656 curator-claim queries spanning 862 unique Dictyostelium genes, linked to 1,289 unique PubMed articles (2,028 claim–article pairs), evaluated against a 20,447-article Europe PMC abstract corpus.

Loading

from datasets import load_dataset

# Main benchmark: curator claims with their cited articles (the default config, `goldset`)
queries = load_dataset("fulaibaowang/dictycite", split="train")

# Retrieval corpus
corpus = load_dataset("fulaibaowang/dictycite", "abstract_corpus", split="train")

Each config is a single JSONL file, exposed under the default split name train. The name carries no meaning: DictyCite is an evaluation benchmark and has no train/test partition.

Config	File	Rows	Contents
goldset	dictycite_goldset.jsonl	1,656	Curator claims, cited PubMed articles, LLM evidence-level labels
abstract_corpus	dictycite_abstract_corpus.jsonl	20,447	Cleaned abstracts forming the retrieval corpus
query_expansion	dictycite_query_expansion.jsonl	563	Goldset subset with gene-aware query-expansion variants

The repository also holds dictycite_fulltext_pmids.tsv (1,124 rows): the cited PMIDs that have a chunked full-text representation, with PMC licence tags. It is a plain file, not a config:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("fulaibaowang/dictycite", "dictycite_fulltext_pmids.tsv", repo_type="dataset")
fulltext = pd.read_csv(path, sep="\t", dtype={"pmid": str})

Queries join to the corpus on docs[].pmid = pmid:

pairs = [
    (q["query_id"], d["pmid"], d["evidence_level"])
    for q in queries
    for d in q["docs"]
]

Schemas

goldset

One row per query.

• query_id - string. Stable query identifier.
• query_text - string. Curator-written claim used as the retrieval query.
• documents - list of strings. PubMed URLs of the cited articles.
• docs - list of objects. Cited articles with metadata (below).
• genes - list of objects. dictyBase gene records linked to the query (gene_id, gene_name, synonyms, gene_products).
• query_gene_expansion - object. Gene-detection result (has_detectable_gene, n_detected_genes, detected_gene_ids, detected_gene_expandable, query_expansion_benchmark).

Each entry in docs:

• pmid - string. PubMed ID (joins to abstract_corpus.pmid).
• publication_id - int. dictyBase internal publication ID.
• title - string. Article title.
• year - int. Publication year.
• abstract_clean - string. Cleaned abstract text.
• evidence_level - string. LLM label, one of abstract_supports_detail, abstract_supports_core or abstract_insufficient (844 / 762 / 422 of the 2,028 pairs).
• doc_match - string. yes if the cited article was matched to the abstract corpus.
• reason - string. LLM rationale for the label (may be empty).
• anchor_pos - list of ints. Character offsets of the citation anchors in the source curator note.
• citation_captions - list of strings. Raw citation strings from the dictyBase page.

The evidence-level labels were assigned by Llama 3.3 under a fixed prompt, not by human annotators. They describe how much of the claim the abstract alone supports.

abstract_corpus

One row per article.

• pmid - string. PubMed ID (primary key).
• docno - string. Indexer document number (equal to pmid).
• pmcid - string or null. PMC ID, when available.
• doi - string or null. DOI, when available.
• title - string. Article title.
• authors - string or null. Comma-separated author list.
• journal - string. Journal name.
• year - string. Publication year.
• text - string. Cleaned abstract text.
• type - string. Always "abstract".
• file - string. Source file path (provenance only).

query_expansion

The 563 goldset queries with query_expansion_benchmark == "yes". Fields as in goldset, except that the gene records sit under query_gene_expansion.detected_genes instead of a top-level genes field, plus two expanded-query fields:

• query_text_expansion_synonyms - string. Original query with gene synonyms appended.
• query_text_synonym_products - string. Original query with gene synonyms and product descriptions appended.

dictycite_fulltext_pmids.tsv

One row per cited PMID that has a chunked full-text representation in the paper's pipeline.

• pmid - PubMed ID (joins to docs[].pmid; read it as a string).
• pmcid - PMC ID, empty when the article is not in PMC.
• license - PMC licence tag (CC BY, CC BY-NC, CC0, NO-CC CODE, ...) or NON-OA when the article is not in the PMC Open Access Subset.
• n_chunks - number of full-text chunks for this PMID.

Full text

The paper also reports results on a chunked full-text corpus. The chunked text is not distributed, here or on Zenodo, because most of the underlying articles are outside the PMC Open Access Subset. dictycite_fulltext_pmids.tsv lists which cited articles were covered and under which licence; the PDF processing code is in the GitHub repository under scripts/public/pdf_processing/.

Licence

The data files are released under Creative Commons Attribution 4.0 International (CC BY 4.0), the same licence as the Zenodo record.

Upstream attribution:

• Curator claims and gene annotations are derived from dictyBase gene summary pages. Please cite dictyBase when using the goldset.
• Abstract text is sourced from PubMed via Europe PMC. NLM does not claim copyright on PubMed abstracts, and abstracts may be subject to publisher or author copyright. Users are responsible for complying with the applicable NLM/PubMed and publisher terms.

Citation

Please cite the paper, and the Zenodo record for the dataset itself:

Yun Wang, Gad Shaulsky, Tomaž Curk, Blaž Zupan. Benchmarking Literature Retrieval for a Model Organism: A Dictyostelium Case Study. Discovery Science (DS 2026), Springer, 2026 (to appear). arXiv:2610.03130.

Yun Wang. DictyCite: A Domain-Specific Biological Literature Retrieval Dataset for Dictyostelium discoideum. Zenodo, 2026. https://doi.org/10.5281/zenodo.20308282

Funding and acknowledgements

This work was funded by the European Union (Horizon Europe MSCA COFUND grant No. 101081355 - SMASH). We thank the dictyBase curators for the gene annotations and curator notes that make this resource possible, and the National Library of Medicine and Europe PMC for open access to PubMed metadata and abstracts.