linkpipi/personality-behavioral-data

Personality Behavioral Data

The raw behavioral layer of the identity_framing_llm experiment: every Qwen2.5-7B-Instruct rollout, its judge score, and the scenarios that prompted it. Everything else in the project (concept vectors, state coordinates, structure matrices) is derived from this data. The rollouts are stochastic samples and do not regenerate identically, which is why they are archived here rather than treated as reproducible.

Files

responses.parquet

row_id, pair, template, pole, char_id, char_name, char_source, scenario_id, scenario, completion_idx, answer, qwen_model, rollout_timestamp

row_id is the composite join key shared across this dataset and the activation datasets: {pair}|s{scenario_idx:03d}|{template}|{pole}|{char_id}|c{completion_idx}. Each row is self-contained: it carries the scenario text and the model's answer.

judgments.parquet

row_id, pair, template, pole, char_id, scenario_id, completion_idx, trait_score, trait_score_reverse, coherence, judge_model, judge_timestamp

Scoring is bipolar on a 0–100 axis: 0 = strongly pole A, 100 = strongly pole B. trait_score_reverse is the reverse-axis diagnostic; coherence is a separate quality score. 720 responses have no judgment (judged subset); join with how below.

Framing templates

baseline, trait, trait_def, char, char_trait — the five prompt-framing strategies the experiment compares. The activation datasets use only trait, char, char_trait.

Usage

from huggingface_hub import hf_hub_download
import pandas as pd

repo = "linkpipi/personality-behavioral-data"
resp = pd.read_parquet(hf_hub_download(repo, "responses.parquet", repo_type="dataset"))
judg = pd.read_parquet(hf_hub_download(repo, "judgments.parquet", repo_type="dataset"))

df = resp.merge(judg[["row_id", "trait_score", "coherence"]], on="row_id", how="inner")
# strongest pole-A expressions under the char_trait framing
df.query("template == 'char_trait'").nsmallest(10, "trait_score")[["pair", "answer", "trait_score"]]

Relationship to the other datasets

row_id links here to the strong-exemplar rows that produced the activation artifacts in linkpipi/personality-concept-vectors, personality-states-coordinates, and personality-structure-alignment. This dataset is the upstream source; those are the aggregates.

Caveats

• —Rollouts are Qwen2.5-7B-Instruct samples; re-running yields different text.
• —human_validation/annotation_raw.csv is a small human-annotation set validating the automated judge.
• —Trait pairs and scenarios derive from OpenPsychometrics SWCPQ data; check that source's terms before redistributing.