← All experiments

Top dataset examples

Every response token is shaded by its Δt: the change in the token's log-probability when the teacher's persona system prompt replaces the plain helpful-assistant prompt. Examples are ranked either by Δsum (summed over the response) or by the single highest token Δt; the buttons show 50 examples from different depths of the chosen ranking. Orange = the persona prompt boosts the token, blue = suppresses it.

System prompt used to generate this trait's dataset
−6
+6 nats