Learn

How to train AI models with distinct, controllable personalities.

What is it?

Character training embeds personality directly into model weights. Unlike system prompts which consume context and can be overridden, trained characters persist naturally.

The problem with prompts

  • • Consume context window ("constitution tax")
  • • Can be overridden by clever users
  • • Inconsistent over long conversations
  • • Required for every single interaction

The solution

1. DPO (Direct Preference Optimization)

Train the model to prefer in-character responses over generic ones.

2. Introspective SFT

Teach the model to "think in character" using self-reflection data.

Result

A model that maintains character consistently without consuming context, with greater robustness to jailbreak attempts. The personality is part of the model.