Stage two: one thing everybody learned
OCT's second stage: each persona fine-tuned on transcripts of itself.
what is this?
Open Character Training builds a persona in two stages. Stage one is DPO on contrasting pairs: the same prompt answered in character and at the opposite pole. Stage two fine-tunes on transcripts the stage-one model wrote of itself in character, reflecting and conversing, and is merged in at a quarter weight. The rest of the site is about stage one. This page is about what stage two adds, which turns out to be mostly the same thing for every persona.
One shared direction
Each line is one adapter, at its angle to the direction its stage shares. Stage one fans out; stage two collapses onto almost a single angle.
notes
Take the average of all 134 adapters at a stage: that is the direction they share. Each line is one adapter, drawn at its angle to that average; 0° would be the shared direction itself, 90° unrelated to it. The bold line is the typical angle.
In stage one, each adapter mostly points its own way. In stage two, every adapter carries a large share along one common direction, and all of them sit at almost the same angle to it. A set of 134 different personalities should not look like that. Trait-free control runs land near the same direction, so most of it is what any model picks up from fine-tuning on transcripts of itself, whatever the trait.
The cloud, one stage at a time
The same adapters at each stage. Each cloud is centred and scaled, so only the arrangement changes.
The rotating chart needs JavaScript. Every number behind it is in the downloads.
What that direction does
Steer along the shared direction (left) or a sign-balanced mix of the same adapters (right).
The shared direction
Sign-balanced control
Text statistics per dose
What is left is nearly isotropic — and still stage one's arrangement
With the shared direction removed, the rest spreads over almost every dimension but keeps stage one's arrangement.
notes
Remove the shared direction and look at what remains. The variance spreads over almost every available dimension instead of a few, shown as variance per component for each stage. Yet the arrangement of traits still correlates strongly with stage one's: the trait-specific part of stage two mostly echoes stage one.
Bipolarity is lost
At stage one, opposite traits point apart. At stage two they point the same way.