Bibliometric Dataset: Defense Mechanisms in Large Language Models, 2024–2026

Reproducibility package for the bibliometric mapping of defense mechanisms against adversarial attacks on Large Language Models (LLMs), accompanying the manuscript "The Evolution of Defense Mechanisms in Large Language Models: Emerging Trends and Strategic Mapping of the 2024–2026 Research Landscape" submitted to Scientometrics. This deposit contains: Raw corpus: a Scopus BibTeX export of 137 Q1 journal articles published 2024–2026, extracted on 13 February 2026 from 17 explicitly white-listed Q1 sources (SCImago Journal Rank 2024). 2. Search queries: the verbatim Boolean queries used in Scopus, including the main corpus query and the methodological precedent query. 3. Analysis code: a fully reproducible R script using Bibliometrix 4.1 that regenerates every quantitative claim in the manuscript — main bibliometric indicators, annual scientific production, Bradford's Law, Lotka's Law, country productivity and collaboration networks, author co-authorship networks, most cited documents, Callon's strategic centrality–density diagram, thematic evolution analysis, Multiple Correspondence Analysis (MCA), hierarchical clustering, and keyword co-occurrence networks. 4. Pre-computed results: 13 Excel tables exported directly from Bibliometrix on 13 February 2026 (the actual files that produced the reported values), plus the country-level production CSV and the preliminary knowledge-gap analysis generated with Scopus AI Deep Research. 5. Visualizations: 23 figures used in or supporting the manuscript, including VOSviewer overlay, density, and network visualizations. 6. Documentation: a comprehensive README, data dictionary, search strategy document, citation file (CITATION.cff), and CC BY 4.0 license. Key reproducible findings: 137 documents from 19 unique Q1 sources, 78.38% annual growth rate, 516 authors, China–Singapore as the most active bilateral collaboration dyad (15 links), and "adversarial machine learning" as the highest-frequency motor theme (n = 23 keyword occurrences). Multiple Correspondence Analysis explains 53.40% of conceptual variance across two dimensions. Software requirements: R ≥ 4.3, Bibliometrix 4.1.4, igraph 2.0+, VOSviewer 1.6.20. Suggested citation: Vargas Yáñez, S., & Tobón, S. (2026). Bibliometric Dataset: Defense Mechanisms in Large Language Models, 2024–2026 [Dataset]. Zenodo.