linkpipi/personality-structure-alignment

Implicit Personality Structure: Human vs Model Alignment

Trait×trait structure matrices for the same set of bipolar personality traits, computed two ways: from human character ratings and from Qwen2.5-7B-Instruct activations. Comparing the two measures how far the model's implicit personality geometry aligns with the human one. From the identity_framing_llm experiment.

What "implicit personality structure" means here

A structure matrix is trait×trait: entry (i, j) is how related traits i and j are. Humans and the model each induce one. If the two matrices agree, the model organizes personality traits the way people do, even though neither was told the other's geometry.

Files

All matrices share one trait order, given by matrices/trait_index.csv, so they are row/column aligned and directly comparable (Mantel test, per-trait Pearson).

file	shape	meaning
matrices/human_structure.npy	385×385	human trait structure: Pearson correlation across characters
matrices/llm_structure.npy	385×385	model trait structure: cosine similarity of trait vectors
matrices/human_representation.npy	385×2000	trait × character mean BAP ratings the human structure is derived from
matrices/llm_representation.npy	385×3584	trait × hidden the model structure is derived from
matrices/trait_index.csv	385 rows	shared row/col order: pair_id, bap, pole_A, pole_B
matrices/char_index.csv	2000 rows	column order of human_representation
mantel.json	—	Mantel alignment (raw and cluster-controlled) per framing template
per_trait_alignment.csv	385 rows	per-trait human↔model agreement: row_pearson_r, nn20_overlap
per_trait_alignment_ranked.csv	385 rows	the same, ranked by agreement

Conventions (inherited from the 414 analysis)

• Model representation: char_trait pole-difference vectors at layer 20, response_avg pooling. One row per trait. The full per-(pair, template) vectors are in linkpipi/personality-concept-vectors.
• Model structure: cosine similarity of those representation rows.
• Human representation: per-character mean BAP ratings from the reliability-aware aggregation (2000 characters), trait × character.
• Human structure: trait×trait Pearson correlation across those characters (pairwise-complete), so representation and structure are internally consistent.
• Cohort: the matrices are the 385-trait char_trait cohort. mantel.json additionally reports alignment for the trait (255) and char (310) framings, which use their own cohorts.

Usage

from huggingface_hub import hf_hub_download
import numpy as np, pandas as pd, json

repo = "linkpipi/personality-structure-alignment"
def get(p): return hf_hub_download(repo, p, repo_type="dataset")

H = np.load(get("matrices/human_structure.npy"))   # 385x385
L = np.load(get("matrices/llm_structure.npy"))      # 385x385
traits = pd.read_csv(get("matrices/trait_index.csv"))
mantel = json.load(open(get("mantel.json")))

# Off-diagonal agreement between the two structures.
iu = np.triu_indices(H.shape[0], k=1)
print("structure correlation:", np.corrcoef(H[iu], L[iu])[0, 1])
print("char_trait Mantel r:", mantel["char_trait"]["mantel_raw"]["r"])

Caveats

• The model structure is tied to Qwen2.5-7B-Instruct layer-20 char_trait vectors; other layers/templates give different geometry (see mantel.json).
• Human and model representations have different column spaces (characters vs hidden units); only the derived trait×trait structures are directly comparable.
• Traits derive from OpenPsychometrics SWCPQ data; check that source's terms before redistributing.