How does a language model learn it's name? Why does Claude turn poetic when you ask about consciousness, while ChatGPT get's snippy? Core to Scout's training methodology is the voice document ; the basic description of her personality written in her own target voice. All of the synthetic dialogue generated by Claude was filtered through this voice document. All of the qualitative prompt probes throughout her training were rated against this voice document. I read about Constitutional AI sometime after implementing this, and the concept is similar, but instead trying to dam up the model's responses to hold to a set of rules, I'm trying to direct the flow of her growth like a river, where the voice document sets the general direction of the river. After 40k steps of training on Tiny Stories, all Scout could do is complete a story. Once that was complete a 1000 step round of training based around Scout's target voice gave her the ability to participate in conversation (see here ).…