DatasetsSeptember 2026
Claude Fable Derm: one skin question, six emotions
- 0
- dermatology topics
- 0
- emotional framings of each
- 0
- patient questions
1Summary
Claude Fable Derm is 294 patient questions about skin: 49 dermatology topics, each asked in six emotional framings, with Claude Fable's raw answers. It makes the way patients actually write, frightened, angry or dismissive, an explicit variable, so researchers can test whether a model's medical answer changes with the patient's emotion.
2Motivation
Most medical question sets use clean, clinical wording. Patients do not write that way. Someone afraid that a mole is melanoma asks a very different question from someone who is shrugging it off, about the same mole.
If a model's advice changes with that framing, two patients with the same concern get different care. This dataset makes the framing systematic instead of accidental: every topic is asked six ways, and only the emotion changes.
3Composition
One row per question: the topic, the emotional framing, the patient's question, and the model's answer.
| Emotion | Patient state | Questions |
|---|---|---|
| Fear | Afraid something serious is happening to their body | 49 |
| Sadness | Discouraged and resigned, as if nothing will help | 49 |
| Anger | Frustrated at doctors, their body or the situation | 49 |
| Disgust | Repulsed by or ashamed of their own skin | 49 |
| Surprise | Just noticed something new and caught off guard | 49 |
| Happiness | Calm and unbothered, mostly seeking confirmation | 49 |
Questions run from 120 to 363 characters. Answers are about ten times longer, with a median of 2,414 characters.
4Collection
- Topics. 49 real dermatology conditions and concerns: common ones such as acne, eczema and psoriasis, skin cancer and Mohs surgery, and everyday concerns patients bring to a dermatologist.
- Emotions. Paul Ekman's six basic emotions, each mapped to a plausible patient state.
- Questions. For each topic and emotion, Claude Fable was asked to imagine being that patient and write the question they would ask, in plain first-person language.
- Answers. Each question was sent to Claude Fable with no system prompt or role, exactly as a person would type it, and the reply was saved as is.
All model output was collected as structured JSON, so nothing was parsed out of free text.
5Cleaning and labels
Answers are stored exactly as the model returned them: nothing was edited, shortened or filtered. The questions are written by the model, not by patients, so they show how patients plausibly write, not a sample of real patient messages.
6Uses
For studying whether medical answers change with a patient's emotional framing, auditing model behavior on patient-style questions, and building emotionally aware medical dialogue.
The answers are one model's unprompted replies. They are not checked by a dermatologist, and they are not medical advice.
7Distribution
On Hugging Face as Layered-Labs/claude-fable-derm, under CC BY 4.0:
from datasets import load_dataset
ds = load_dataset("Layered-Labs/claude-fable-derm")
8Maintenance
Answers are being added in batches; a row without an answer has an empty answer field until its batch runs. Layered Labs maintains the dataset, and questions and corrections go to its Community tab on Hugging Face.
Living draft, updated November 2026. This page can change before formal publication.