MASLD-LLM: analysis plans, code, and benchmark results for LLM vs classical-ML fibrosis staging from cfRNA gene-expression prompts

Benchmark package for the study "Can large language models read cell-free RNA expression panels for fibrosis staging in MASLD?" Twelve frontier LLMs were scored against classical machine-learning gates (logistic regression, HistGradientBoosting) on three task families — binary advanced fibrosis, 5-class staging, and ordinal 0–4 — using text-formatted cfRNA prompts derived from the top-200 MAD genes of the Ghent plasma cell-free transcriptome cohort (Poma-Soto et al.). This version contains the Colab notebooks (prompt construction, baseline and frontier evaluation, automated run drivers, C3 evaluations, v6 fine-tune and evaluation), the static benchmark site (task catalogue, methods, models, leaderboards with bootstrap confidence intervals), one summary figure, and a frontier-model run log. It does NOT contain per-sample gene-expression values or the full prompt texts, which derive from EGA-gated data (EGAD50000000775, Ghent DAC Data Usage Agreement). The archived analysis plans, run/deviation log, and corpus manifest are not included in this version.Acknowledgements The Ghent cohort was generated by the Ghent University research groups responsible for the plasma cell-free RNA biobank and cohort, who act as the Ghent DAC for the EGA deposition (EGAD50000000775 / EGAS50000000545). Subhadip Banerjee acknowledges the Research Foundation – Flanders (FWO), which allocated a grant for a scientific stay in Flanders (file ID V504225N, hosted at the University of Antwerp). Rawiwan Charoensub acknowledges the Ursimone Wietlisbach Foundation for grant support.