Fixed word lists lose track of large language model use in biomedical writing: data and code

Data and code for the manuscript "Fixed word lists lose track of large language model use in biomedical writing" (Department of Trauma Surgery, University Hospital Regensburg). Version 2.0 extends version 1 to September 2026 and adds every analysis of the present manuscript. The record contains three archives. deposit_nhb.zip holds the base data and code for 2015 to 2025: the eligible identifier lists for 77 sampling strata (1,113,354 open-access records), the marker word counts for 41,403 stored full texts (37,996 distinct texts from 37,691 articles after removal of 3,407 duplicates), the marker word list with its pre-2022 frequencies and the retrieval and analysis notebooks; it is described in its own README.md, and its peer review and specialty analyses are not used in the present manuscript. code_2026_extension.zip holds the code for every analysis added for the present manuscript, in execution order, and the new word lists. data_2026_extension.zip holds the identifier lists, word counts and output files of the 2026 extension. All code ran in Google Colab (Python 3.12). README_2026_extension.md gives the execution order and the output file of every script, and Supplementary Table 8 of the manuscript maps every reported estimate to its output file. Full texts and abstracts are third-party content and are not redistributed. Full texts can be retrieved again from the deposited PMCIDs through the Europe PMC REST API, abstracts from the deposited PMIDs through the NCBI E-utilities. The published marker list originates from berenslab/llm-excess-vocab (MIT licence; Kobak et al., Sci. Adv. 11, eadt3813, 2025). Data are licensed under CC BY 4.0, code under the MIT licence.