I fine-tuned an LLM to be C-3PO to test which training data format works best for persona injection [P]

In three linesLoRA fine-tuning experiment comparing three data formats for C-3PO persona injection: chat demos, first-person statements, and synthetic Wikipedia docs. First-person statements win on generalization. Synthetic docs produce paradoxical behavior: model knows C-3PO is anxious but expresses it only 37% of the time.

Summary generated by Claude — human-verified