Notebook 27 — Train SAE on a fine-tuned model (LoRA)

[caiovicentino]

User has a LoRA-fine-tuned checkpoint, wants to train an SAE on its residuals. Document the PEFT hook path + warn about LoRA merging. Target Gemma-2-2B + a small LoRA.