ESR-AbstractRater: European Radiology & Diagnostics article corpus and local LLM recommendation outputs (2026)

This dataset accompanies the study Can Local Large Language Models Discriminate Journal Selectivity? A Comparative Evaluation Using Published Radiology Articles (F1000Research, 2026). It contains (i) the PubMed record list of the 2,196 articles screened and (ii) all 13,008 recommendations produced by four locally deployed large language models, evaluated against two target journals (European Radiology and Diagnostics) under an abstract-only and a full-text condition. Files articles.csv (2,196 rows): PubMed metadata and abstracts. Columns: pmid, doi, title, authors, journal, journal_iso, journal_query, source_journal, publication_date, publication_types, abstract, url_pubmed. The 813 records with an abstract constitute the analysed corpus (European Radiology 527, Diagnostics 286). inference_results.csv (13,008 rows): model recommendations. Columns: model, condition, pmid, doi, source_journal, target_journal, recommendation, confidence, json_valid, duration_ms, raw_response. Join to articles.csv on pmid. README_data_dictionary.md: full data dictionary, PubMed query strings, processing steps and scope of the licence. European Radiology serves as the higher-selectivity target journal; Diagnostics serves as a silver-standard, lower-selectivity comparator. Full texts are not redistributed: they were retrieved under institutional subscription and remain under publisher copyright. Each full text can be re-acquired from the DOI given for the corresponding article; the acquisition code is available in the extended data record. Abstracts are reproduced from PubMed for research and reproducibility purposes; copyright in the abstracts remains with the respective publishers. The CC-BY 4.0 licence applies to the compilation and to the authors' own contributions (record selection, annotation, model outputs and analyses).