Alignment

Narrow Fine-Tuning, Broad Misalignment in a 3B Model

I fine-tuned a 3.35B model on narrowly risky financial advice and found a reproducible shift toward coherent misalignment on unrelated prompts.