DictyCite: A Domain-Specific Biological Literature Retrieval Dataset for Dictyostelium discoideum

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 inline-cited PubMed articles. The release contains: 1,656 curator-claim queries spanning 862 unique Dictyostelium genes, linked to 1,289 unique PubMed articles. A 20,447-article Europe PMC abstract corpus used as the retrieval corpus for all abstract-only experiments in the paper. A 563-query query-expansion subset with two gene-aware expansion variants (synonyms; synonyms and product descriptions) derived from dictyBase gene annotations. LLM-assigned evidence-level labels (abstract_supports_detail, abstract_supports_core, abstract_insufficient) for every claim–article pair, produced with Llama 3.3 under a fixed prompt schema. A list of 1,124 cited PMIDs that have a chunked full-text representation in our pipeline, with PMC Open Access license tags. The chunked text itself is not redistributed here because the underlying PDFs include articles outside the PMC OA Subset; the reproduction code is in the GitHub repository. See the README for file schemas, a loading example, and license details. Code, pipeline configurations, and reproduction instructions are at DictyCite GitHub repository