Datasetpublic

build-small-hackathon/chhaya-skin-extract

Chhaya Skin-Extract Fine-tuning data for Chhaya — a skin & heat-health companion for outdoor workers. Each example is image + "skin check" → findings JSON, teaching MedGemma-1.5-4B to emit Chhaya's structured schema directly (no chain-of-thought preamble) with a concern level grounded in real clinical labels. Why two sources ISIC-2024 SCIN Image type Curated dermatologic close-ups Real consumer phone photos concern ground truth Biopsy diagnosis… See the full description on the dataset page: https://huggingface.co/datasets/build-small-hackathon/chhaya-skin-extract.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

0likes12downloads

Dataset Card

Chhaya Skin-Extract

Fine-tuning data for Chhaya — a skin & heat-health companion for outdoor workers. Each example is image + "skin check" → findings JSON, teaching MedGemma-1.5-4B to emit Chhaya's structured schema directly (no chain-of-thought preamble) with a concern level grounded in real clinical labels.

Why two sources

	ISIC-2024	SCIN
Image type	Curated dermatologic close-ups	Real consumer phone photos
`concern` ground truth	Biopsy diagnosis (Benign/Malignant/Indeterminate)	Dermatologist labels (top-1, confidence ≥ 3)
Strength	Pathology truth, melanoma signal	Real-world robustness, skin-tone diversity

Combining them means the model learns both "what a clinician confirmed this is" and "what a phone photo in daylight actually looks like."

Composition

—1,406 examples — train 1,265 / test 141
—Concern balance: low 700 · see_doctor 366 · watch 340 (`low` downsampled to fight under-warning bias — the one mistake a skin-screening tool must not make)
—Source: ISIC 902 · SCIN 504
—Skin tone (Monk scale, SCIN rows): spans 1–7, concentrated at 2–4 (deliberately includes darker skin tones under-represented in dermoscopy sets)

`concern` mapping policy

Auditable in concern_map.py. Cautious by design:

—see_doctor — malignant / pre-malignant (melanoma, BCC, SCC, actinic keratosis) or urgent (herpes zoster, cellulitis, vasculitis)
—watch — ambiguous pigmented / proliferative (nevus, seborrheic keratosis) or persistent inflammatory (psoriasis, lichen planus)
—low — ordinary / clearly benign (eczema, acne, tinea, contact dermatitis)
—safety override: any low whose description shows the full ABCDE high-risk combo (asymmetry + irregular border + uneven colour) is bumped to watch.

Sun/heat-driven conditions (miliaria, photodermatitis, actinic keratosis) also populate heat_sun_signals, tying the model to Chhaya's purpose.

Schema

json

{
  "what_i_see": "...", "spot": {"type","color","borders","symmetry","texture"},
  "heat_sun_signals": [], "concern": "low|watch|see_doctor",
  "concern_reason": "...", "image_quality": "good|limited", "summary": "..."
}

Provenance & licence