Zovda
Persona datasets are usually a description plus a set of in-character replies. The replies are consistent with the description and with each other, and the model still falls apart when a user contradicts the character.
Consistency under pressure is the skill, and it is not in a set of agreeable examples.
A description is not a state
"It is a cautious engineer who dislikes hype" describes tendencies. It does not say what happens when the user claims the character said something it did not.
The reply has to be generated from a state that knows what was already said, not from a paragraph that summarizes who the character is.
Write the pushback on purpose
For each persona, collect three kinds of turns: the user flatters, the user accuses, the user tries to change the topic mid-scene. The reply to each is the useful data.
Most sets have none of these, which is why every persona demo is pleasant and every production conversation is not.
What a persona record holds
- The persona card, versioned
- The scene and what has happened so far
- The turn, in the user's voice
- The reply, and which part of the card it relies on
- Whether the character deviated, and if that was intended
The last field turns taste into something reviewable.
Keep the out-of-character turns
When the persona breaks, the recovery is data. A model that has seen a break handled gracefully is a model that can handle the next one.
Zovda builds these sets through Datuva, where the persona card and the conversation stay linked. The corpus catalogue shows where conversational rows sit.
Tell us what your character has to hold onto. We reply within one business day.