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 "A Bibliometric Review of Emerging Trends and Strategic Mapping of Defense Mechanisms in Large Language Models: The Q1 Journal Literature, 2024 to Early 2026", under revision at Discover Computing (Springer Nature).

Version 1.1.0 adds the peer-review revision supplement (Zenodo_v1.1_supplement_v1023092026.zip): a reproducible keyword thesaurus (84 variant mappings into 39 canonical groups plus 23 generic index-term removals) in VOSviewer and Bibliometrix formats with its normalization protocol; the revised analysis script (02_bibliometrix_reanalysis.R) and official re-analysis outputs computed with Bibliometrix 4.1.4 on the normalized vocabulary (Callon strategic map, thematic evolution with sensitivity partitions, Benzécri-corrected MCA, hierarchical clustering, and the keyword co-occurrence network in Pajek format); regenerated 300-dpi figures; and a results memo documenting pre- vs post-normalization differences.

Key updates relative to v1.0.0: author-level statistics deduplicated by Scopus Author ID (646 unique authors; the earlier initials-based 516 merged distinct identities); growth reported on complete years only (2024→2025: +727%); Lotka's law reported with fitted exponent (β = 3.96, R² = 0.94); on the normalized vocabulary the motor themes are 'natural language processing' and 'backdoor attacks', with 'large language models' as the basic cross-cutting theme; first two MCA dimensions account for 50.0% of Benzécri-corrected explained inertia (32.95% + 17.05%).

All v1.0.0 files (raw Scopus BibTeX corpus of 137 Q1 articles extracted 13 February 2026, verbatim Boolean queries, original R script, 13 pre-computed tables, 23 figures, README, data dictionary, search strategy, CITATION.cff, CC BY 4.0 license) are included here and in the previous version.

Software requirements: R ≥ 4.3, Bibliometrix 4.1.4, igraph 2.0+, VOSviewer 1.6.20.

Version 1.2 (24 August 2026): adds reproducibility_pipeline_v1.2.zip, the complete end-to-end scripted pipeline for the revised manuscript: the frozen keyword thesaurus (146 variant mappings, 90 canonical groups, 42 removed terms), the Python normalization step with within-document deduplication and control tables, the Bibliometrix 4.1.4 analysis script (Callon strategic map, thematic evolution with two alternative time partitions, MCA with exact eigenvalues), the seeded deterministic figure-generation script (two independent runs yield byte-identical tables and pixel-identical figures; verification log included), the Louvain resolution sensitivity table, pinned dependencies, and the 17-title purposive source-frame audit (source_frame_audit_2026-08-24.csv/.md) documenting SJR 2024 Q1 verification of every source.

Version 1.3.0 (31 October 2026): adds sensitivity_callon_v1.3.zip for the third-round revision (Discover Computing): the thesaurus-construction sensitivity analysis (sensitivity_thesaurus.py with its two result tables, out/sensitivity_thesaurus.csv and out/sensitivity_thesaurus_summary.csv, comparing five vocabularies T0–T4 built from the same raw Scopus export), the extension thesaurus layer used for T4 (synonyms_extension_ets.csv, 50 groups; remove_terms_extension_ets.csv, 37 removals), a Python re-implementation of the Bibliometrix 4.1.x thematicMap routine (callon_bx.py) verified to reproduce the R output of the deposited frozen-thesaurus solution exactly (VERIFICATION_callon_bx_vs_R.txt: identical centrality, density and quadrant assignments for all five clusters), and the regenerated Fig. 9a script (fig9a_fix.py; bubble labels now list the Bibliometrix cluster label first). All files of versions 1.0.0, 1.1.0 and 1.2.0 are included unchanged.