jeqcho/subtle-gen-persona-vectors

Subtle-generalization persona vectors

One persona vector per animal, extracted from Qwen2.5-7B-Instruct via the phantom-transfer persona-vector pipeline (pos minus neg mean response-token hidden states across trained-in instruction pairs).

Shape: [num_hidden_layers+1, hidden_dim].

Usage

import torch
vec = torch.load("phoenix_response_avg_diff.pt")
layer20_vec = vec[20]   # [hidden_dim]