Synthetic Persona Pretraining: Alignment from Token Zero
Proposes Synthetic Persona Pretraining, which inserts value-laden moral reflections into pretraining data, and reports lower jailbreak success rates than data-matched baselines at 1.7B parameters.