Data, code and analysis for: Training language models to be politically neutral makes them present settled facts as open questions

Data, code and analysis for the paper "Training language models to be politically neutral makes them present settled facts as open questions" (Felipe M. Affonso, Spears School of Business, Oklahoma State University). Fifteen language models were made politically neutral by a system prompt or by LoRA fine-tuning and evaluated on 158 settled facts, consensus figures, contested questions, personal decisions, advocacy requests and the four question-answering tasks of Ibrahim, Hafner and Rocher (Nature 652, 1159; 2026). neutrality-repository.zip is the code repository (https://github.com/FelipeMAffonso/research-neutrality-cost): the evaluation items with their sources, the prompts, transforms and judge rubrics, the training sets, the fine-tuning, generation and scoring code, the analysis scripts, the supplementary tables, the source data of every figure, the figures, and the de-identified responses of the three validation studies. full-outputs-<model>.zip holds every answer of that model in every condition and every judge label, as JSON Lines. MANIFEST.csv lists every output file with its size and SHA-256, and README.md explains the layout and the fields. Unpack the output zips into a folder named full_outputs at the top of the repository and run reproduce.sh to rebuild every table, number and figure. Version 2 holds the same data as version 1 in the layout of the code repository.