Data and code for: Does asking in Turkish change self-triage advice from large language models? A paired Turkish–English vignette study

Data and code package for a paired evaluation of ten large language models on 60 physician-written self-triage vignettes (30 scenarios), presented in English and in an LLM-translated, author-checked Turkish adaptation, three times each (3,600 responses), with an exploratory arm of 1,800 multi-turn conversations with a simulated patient. The package contains the vignette sets and back-translations, prompt templates, raw run records and parsed responses, run and analysis code (Python, R), canonical analysis outputs, independent re-computation scripts, the frozen statistical analysis plan and the deviation record. Raw Command A+ (Cohere) responses are not included because the provider's terms restrict benchmarking use and publication of outputs; they are available from the corresponding author on request, subject to Cohere's terms of use. Data: CC BY 4.0; code: MIT. Source vignettes: Ramaswamy et al. (doi:10.5281/zenodo.18451491), MIT licence. See README.md for reproduction steps.