Persona Geology
One LoRA per trait adjective, 134 of them, trained on Qwen3.5-4B. Their weight updates sort into five factors.
what is this?
We trained one small adapter (a LoRA) on Qwen3.5-4B for each of 134 personality adjectives: Goldberg's 100 Big Five markers plus 34 words held out from a wider trait lexicon. Every adapter used the same recipe, prompts and starting point; only the trait changed. We then measured how similar every pair of weight updates is and factored that similarity matrix, the way psychologists factored people's self-ratings to get the Big Five.
This site lets you look at the result yourself: where each adapter sits, what the factors are made of, what steering along them does to the model's writing, and what the second stage of character training adds. The post tells the story; this is the reference.
The factor chart
Each point is one adapter, placed in the span of the five factors. Drag to turn it, click a point to pin it.
The rotating chart needs JavaScript. Every number behind it is in the downloads.
The five factors
Each factor runs between two poles. These are the traits that load most strongly at each end; the names are our reading of them.
Explore
- [ chart ]Every adapter in three dimensions, what defines each factor, and how many factors to keep.
- [ planes ]Any two factors, flat, with each word group's two poles joined by an arrow.
- [ behaviour ]Steer the model along a factor and read what it writes, beside the judge's scores.
- [ stage two ]What the introspection stage adds: one direction every persona shares.
- [ traits ]One page per adapter: constitution, training pairs, generations.
- [ data ]Every table behind the site, as CSV and JSON.
- [ wiki ]The full record, written for agents: every number with its source.