PersonaBench Aiko

A controlled LoRA experiment testing whether fine-tuning can make a compact language model maintain a consistent original character across held-out prompts.

What the project proves

  • 1.5B parameter language model
  • LoRA fine-tuning
  • Twenty held-out evaluation prompts
  • Prompting baseline comparison
  • Public model, dataset, source code, and character chat

Back to all selected work