danielfdias98/derm-reasoning-label-only

Dermatology Reasoning Dataset — Label Only

This is the label_only ablation variant of danielfdias98/derm-reasoning, republished as a standalone repo so it can be plugged directly into a trainer without picking a config.

The assistant turn for every row contains: diagnosis + clinical category labels only — the ablation baseline that tests whether the reasoning supervision in the full-reasoning variant actually helps.

Quick load

from datasets import load_dataset
ds = load_dataset("danielfdias98/derm-reasoning-label-only")
print(ds)  # DatasetDict {train: 25_637, val: 2_849}

Sister repo

The complementary ablation variant is at danielfdias98/derm-reasoning-full-reasoning. Same train/val split (95/5, seed=42), same images, just different assistant-turn density.

Image sources

The combined parent repo has the full per-source attribution table and image-redistribution split (public-redistributable + private-full image sets). See danielfdias98/derm-reasoning for the canonical README, citations, and download_images.py script.

License

CC-BY-NC-SA 4.0 (matches the most restrictive component of the source datasets, SkinCAP). Non-commercial research use only.

Citation

@misc{dias2026derm-reasoning,
  author = {Ferreira Dias, Daniel},
  title  = {Dermatology Reasoning Dataset: Structured chain-of-thought annotations across five public sources},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/danielfdias98/derm-reasoning}},
}