danielfdias98/derm-reasoning-label-only

---
language: [en]
license: cc-by-nc-sa-4.0
task_categories: [image-text-to-text, visual-question-answering]
size_categories: [10K<n<100K]
tags: [medical, dermatology, vision-language, vlm-finetune, fairness]
pretty_name: Dermatology Reasoning Dataset — Label Only
---

# Dermatology Reasoning Dataset — Label Only

This is the **label_only** ablation variant of [danielfdias98/derm-reasoning](https://huggingface.co/datasets/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

```python
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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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

```bibtex
@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}},
}
```