Supervised fine-tuning (SFT) — is the mechanism that produces → Finetuning shift (activation-shift metric)

The finetuning that produces a finetuning shift is not, in this paper, some exotic or heavily engineered process -- it is one epoch of rank-32 rsLoRA (Kalajdzievski, 2023) with scaling factor alpha=64 and learning rate 1e-5, run on a single H100 GPU (persona-vectors, §"D.3 Finetuning details", p. 38). That modesty is the point: supervised finetuning's role in producing finetuning shift is not to demonstrate that aggressive or prolonged training corrupts a persona, but that almost any ordinary, resource-light adaptation run does, provided the training data itself carries a directional signal. The same low-rank recipe is applied uniformly across trait-eliciting datasets, EM-like datasets, and benign controls, so differences in the resulting finetuning shift are attributable to what the data teaches, not to how hard the model was trained. This uniformity is what lets finetuning shift function as a clean dependent variable across dozens of dataset comparisons: because rank, epoch count, and learning rate are held fixed everywhere, the only thing varying between a high-shift and a low-shift run is the dataset itself, isolating supervised finetuning's contribution to persona drift from the confound of training-intensity differences that a less standardized experimental design would leave unresolved.