triniborrell/manifold-persona-roles-response

Residual-stream activations for 276 character-role archetypes, extracted from Qwen/Qwen2.5-3B-Instruct.

Each point is one system(role instruction) + user question chat. This cloud is the response-token variant: activations are averaged over the generated assistant response tokens (the Assistant Axis paper's token basis).

Contents

file	shape / rows	description
prompt_avg.npy	(6900, 37, 2048) fp16	mean over response tokens, all layers
prompt_last.npy	same	final response token, all layers
metadata.csv	6900 rows	role, is_default, instruction_idx, question_idx, question, system, text, response
manifest.json	—	model, layer counts, extraction settings

Row i of the arrays corresponds to row i of the metadata.

Extraction settings

{
  "model_name": "Qwen/Qwen2.5-3B-Instruct",
  "n_layers": 37,
  "hidden": 2048,
  "primary_layer": 19,
  "token_basis": "response",
  "max_new_tokens": 128,
  "do_sample": false,
  "n_questions": 5,
  "n_records": 6900,
  "n_roles": 276
}

The analysis layer is primary_layer = 19 (~0.5 depth, matching the Assistant Axis paper). prompt_avg/prompt_last keep those names in both variants so downstream code loads either cloud unchanged; token_basis in the manifest records which it is.

Loading

import numpy as np, pandas as pd
from huggingface_hub import snapshot_download

p = snapshot_download("triniborrell/manifold-persona-roles-response", repo_type="dataset")
X = np.load(f"{p}/prompt_avg.npy", mmap_mode="r")   # [N, n_layers, hidden]
meta = pd.read_csv(f"{p}/metadata.csv")
layer = 19
points = np.asarray(X[:, layer, :], dtype=np.float32)

Produced by manifold-persona extraction/.