Reproducible LoRA useful-rank diagnostics from activation-whitened early-gradient spectra, with theory, controlled simulations, GPT-2/WikiText-2 validation, and Zenodo data artifacts.
transformers lora random-matrix-theory peft gpt-2 low-rank-adaptation wikitext-2 effective-rank reproducible-ml rank-allocation activation-whitening gradient-spectra
-
Updated
Jun 30, 2026 - Python