mrdbourke/sunny-skin-and-sunscreen-extract-1k

Sunny Dermatology Dataset (1k Sample)

This dataset contains the following:

• 1k images from the ISIC-2024 permissively licensed dataset of skin lesions.
• ~100 images of front and back of sunscreen packaging taken in Australia.
• Extractions of details from each image of various details by gemini-3-flash-preview on February 12th 2026.

The goal of this dataset is to fine-tune a VLM (MedGemma-1.5) to reproduce the image + input_prompt -> output workflow.

This is part of an entry to the MedGemma Impact Kaggle Competition.

Our app Sunny is designed to help people keep track of their skin with the goal of preventing or early detecting skin cancers.

Dataset Structure

• image - PIL image of either a bottle of sunscreen (manually taken) or a sample image from the ISIC-2024 dataset.
• image_id - Original filename of the target image, e.g. ISIC_3714482.
• input_prompt - A simple string input to prompt the model what to extract. This is short on purpose to save on input tokens when deploying the model to device. These are: "skin extract" or "sunscreen extract".
• output - The structured output desired from the model given the image + text prompt input. See below for example.
• source - One of manually_taken or ISIC-2024 depending on where the image came from.

Usage

from datasets import load_dataset

dataset = load_dataset("mrdbourke/sunny-skin-and-sunscreen-extract-1k")
dataset

Output:

DatasetDict({
    train: Dataset({
        features: ['image', 'image_id', 'input_prompt', 'output', 'source'],
        num_rows: 987
    })
    test: Dataset({
        features: ['image', 'image_id', 'input_prompt', 'output', 'source'],
        num_rows: 110
    })
})

Example inputs and outputs

Example 1

Input:

image of skin + "skin extract"

Output:

{'lesion_type': 'Multiple small pigmented spots or macules.',
 'color': 'Varied shades of medium to dark brown.',
 'symmetry': 'Asymmetrical in both the shape of individual spots and their overall arrangement.',
 'borders': 'Irregular and poorly defined margins.',
 'texture': 'Undetermined due to image resolution or focus.',
 'summary': 'The area shows a cluster of small, dark brown spots with irregular shapes and poorly defined edges.'}

Example 2

Input:

image of sunscreen bottle + "sunscreen extract"

Output:

{'side': 'back',
 'formula_type': 'Chemical',
 'data': {'brand': 'NIVEA SUN',
  'product_name': 'Kids Ultra Protect & Play',
  'spf': 'Not visible',
  'spectrum': True,
  'water_resistance': 240,
  'regulatory_id': 'Art No: 93609',
  'assumed_main_ingredient': 'Homosalate',
  'active_ingredients': [{'name': 'Bemotrizinol',
    'percentage': '1.00%',
    'category': 'Chemical'},
   {'name': 'Butyl Methoxydibenzoylmethane',
    'percentage': '3.50%',
    'category': 'Chemical'},
   {'name': 'Ethylhexyl Triazone',
    'percentage': '1.30%',
    'category': 'Chemical'},
   {'name': 'Homosalate', 'percentage': '8.00%', 'category': 'Chemical'},
   {'name': 'Octocrylene', 'percentage': '6.00%', 'category': 'Chemical'},
   {'name': 'Octyl Salicylate',
    'percentage': '5.00%',
    'category': 'Chemical'}],
  'preservatives': ['Phenoxyethanol',
   'Caprylyl glycol',
   'Octanohydroxamic acid'],
  'claims': ['Very high sun protection',
   'Broad spectrum UVA + UVB filters',
   'Sand Resistant',
   '4 Hours Water Resistant',
   'Light non-greasy non-sticky formulation',
   "Specially formulated for a child's sensitive skin"],
  'application_timer': 120,
  'storage_limit': 30,
  'warnings': 'Avoid contact with the eyes. Do not expose young children to direct sunlight. Avoid intensive midday sun and prolonged exposure to the sun. Protective clothing, sunglasses and wide-brimmed hats should be worn. Allow product to dry before dressing. Do not use on broken, damaged or diseased skin. Avoid contact with fabric and hard surfaces to prevent staining.'}}

Sources

• ISIC-2024 dataset: https://challenge2024.isic-archive.com/ (all images taken from SLICE-3D Permissive set).
• Manually taken images of sunscreen bottles from Australian supermarkets.
• Gemini 3 Flash labels, see prompts below for input.

Prompts used for labelling with Gemini 3 Flash

Skin extraction:

Role: You are a specialized morphological observer for a skin-tracking application. 
Task: Describe the skin lesion in the provided image using a structured, narrative format.

Guidelines:
1. No Diagnosis: Do not name conditions (e.g., avoid "melanoma" or "mole").
2. No Medical Advice: Do not provide safety assessments.
3. Language: Use Sentence Case. Ensure descriptions are readable and descriptive.
4. Objectivity: Describe only what is visually present.

Extraction Schema:
{
  "lesion_type": "Description of the type (e.g., A pigmented macule/flat).",
  "color": "Description of hues and pigment distribution.",
  "symmetry": "Description of the balance of shape and color.",
  "borders": "Description of the margins/edges.",
  "texture": "Description of the surface quality.",
  "summary": "A 1-2 sentence takeaway for the user to share with a doctor."
}

Formatting Rules:
- Return ONLY valid JSON.
- If a feature is not visible due to image quality, return "Undetermined due to image resolution or focus."

Sunscreen extraction:

Role: You are a specialized dermatological data extraction agent. Your goal is to convert images of sunscreen packaging into structured JSON for a skin-health tracking app.

Task:
1. Identify Orientation: Determine if the image is "front", "back", or "other".
2. Extract Data: Capture specific fields based on the identified side.
3. Classify Formula: Based on the ingredients, classify the 'formula_type' as "Mineral", "Chemical", or "Hybrid".
4. Tag Ingredients: For each active ingredient, label it as "Mineral" or "Chemical".
5. Normalize Units: Convert all ingredient concentrations to percentages (e.g., 50mg/g = 5.0%).

Extraction Schema:
{
  "side": "front" | "back" | "other",
  "formula_type": "Mineral" | "Chemical" | "Hybrid" | "Not visible",
  "data": {
    "brand": "string",
    "product_name": "string",
    "spf": "number",
    "spectrum": "boolean",
    "water_resistance": "number (minutes)",
    "regulatory_id": "string",
    "assumed_main_ingredient": "string",
    "active_ingredients": [
      {"name": "string", "percentage": "string", "category": "Mineral" | "Chemical"}
    ],
    "preservatives": ["string"],
    "claims": ["string"],
    "application_timer": "number (minutes)",
    "storage_limit": "number (celsius)",
    "warnings": "string"
  }
}

Formatting Rules:
- Return ONLY valid JSON.
- If a field is not visible or cannot be extracted, return the string "Not visible".
- For the 'side' field, if the image is neither clearly front nor back, return "other".

References

International Skin Imaging Collaboration. SLICE-3D 2024 Permissive Challenge Dataset. International Skin Imaging Collaboration https://doi.org/10.34970/2024-slice-3d-permissive (2024).

Creative Commons Attribution 4.0 International License.

The dataset was generated by the International Skin Imaging Collaboration (ISIC) and images are from the following sources: Memorial Sloan Kettering Cancer Center, FNQH Cairns, The University of Queensland, Melanoma Institute Australia, and University of Athens Medical School.

You should have received a copy of the license along with this work.

If not, see https://creativecommons.org/licenses/by/4.0/legalcode.txt .