Output Projections Matter Most: An Empirical Analysis of Layer-Wise Sensitivity in Time-Series Foundation Model Adaptation

This dataset contains the reproducibility materials for the manuscript “Output Projections Matter Most: An Empirical Analysis of Layer-Wise Sensitivity in Time-Series Foundation Model Adaptation.” It includes Python scripts, study-ready data files, raw experiment logs, summarized CSV results, supplementary experiment outputs, cross-backbone validation results, and figure-generation materials. The experiments investigate projection-level sensitivity, ProjFisher diagnostics, Fisher-guided LoRA rank allocation, implementation-confound controls, and matched-baseline evaluations for time-series foundation model adaptation across CDC ILI, national illness, Solar Energy, PRSA PM2.5, and ETTm2 datasets.

Pre-trained model weights and local model caches are not included; users should obtain the required foundation models from their original providers. The package is intended to support verification of the reported numerical results, matched-control analyses, and manuscript figures.