CoolFace
Datasetpublic

masked-kunsiquat/clinical-personas

clinical-personas Three synthetic clinical personas used as seed data for the Lattice on-device CBT journaling app. Each persona represents a distinct emotional profile on the circumplex model: Persona Entries Quadrant Primary distortions Holmes 35 Q2 (v<0, a≥0) — Tense Catastrophizing, Mind Reading Watson 30 Q3 (v<0, a<0) — Fatigued Emotional Reasoning, Behavioural Avoidance Werther 30 Mixed (high negative valence) Emotional Reasoning, All-or-Nothing… See the full description on the dataset page: https://huggingface.co/datasets/masked-kunsiquat/clinical-personas.

sourceHugging Facemitupdated 6mo agoView on Hugging Face
0likes17downloads
Dataset Card

clinical-personas

Three synthetic clinical personas used as seed data for the Lattice on-device CBT journaling app.

Each persona represents a distinct emotional profile on the circumplex model:

PersonaEntriesQuadrantPrimary distortions
Holmes35Q2 (v<0, a≥0) — TenseCatastrophizing, Mind Reading
Watson30Q3 (v<0, a<0) — FatiguedEmotional Reasoning, Behavioural Avoidance
Werther30Mixed (high negative valence)Emotional Reasoning, All-or-Nothing

Data

The default config (metadata/journal_entries.jsonl) contains one row per journal entry with the following fields:

FieldTypeDescription
idstringUUID, stable across versions
personastringholmes / watson / werther
timestampintUnix epoch milliseconds
contentstring or nullMasked text — real names replaced with [PERSON_<uuid>] tokens
valencefloatCircumplex valence in [-1, 1]
arousalfloatCircumplex arousal in [-1, 1]
moodLabelstringDiscrete label (TENSE, FATIGUED, SERENE, …)
cognitiveDistortionslist[string]Burns distortions identified in the entry
mentionCountintNumber of person mentions in the entry

The activities config (metadata/activities.jsonl) contains Watson's Behavioural Activation activity hierarchy (difficulty 1–10, value category).

Raw seed files (holmes.json, watson.json, werther.json) are also present for direct app ingestion — they include embeddingBase64 (384-dim Arctic Embed XS vectors, IEEE 754 little-endian, base64-encoded) not present in the JSONL.

Usage

python
from datasets import load_dataset

ds = load_dataset("masked-kunsiquat/clinical-personas")
train = ds["train"]  # 95 entries across all three personas

# Filter to a single persona
holmes = train.filter(lambda x: x["persona"] == "holmes")

Privacy

All content fields are pre-masked. No real person names, locations, or identifiers appear in the dataset. [PERSON_<uuid>] tokens are persona-internal references only.