danielfdias98/derm-reasoning-visible-evidence

Dermatology Visible-Evidence Reasoning Dataset

This dataset contains the canonical v2-kept train and validation splits for Visible-Evidence Reasoning supervised fine-tuning. It is derived from the local dermatology reasoning corpus and aligned to the cleaned derm-reasoning-think-v2 rows.

Files

File	Rows	Description
train.jsonl	22,528	Canonical chat SFT training rows aligned to the v2-kept train split.
validation.jsonl	2,539	Canonical chat SFT validation rows aligned to the v2-kept validation split.
metadata/train.v2_kept.audit.json	1	Local filtering audit for the canonical train split.
metadata/validation.audit.json	1	Local validation-generation audit for the validation split.
docs/evidence_reasoning_dataset_implementation.md	1	Project evidence note describing generation, filtering, and validation.

Source Counts

The original fine-tuning corpus contains 25,637 training rows and 2,849 validation rows. The cleaned v2 dataset retained 22,528 training rows and 2,539 validation rows after deterministic quality filtering.

This release uses the v2-kept rows as canonical. A 25,637-row Visible-Evidence Reasoning training superset was generated locally, then filtered back to the v2-kept training image list. The 3,109 excluded train rows correspond to examples previously removed by the v2 audit. The validation split was generated directly from the v2-kept validation file.

Schema

Each row follows a chat SFT structure:

{
  "messages": [
    {
      "role": "user",
      "content": [
        {"type": "image", "image": "final/train/..."},
        {"type": "text", "text": "Diagnose this skin condition with structured visible-evidence reasoning."}
      ]
    },
    {
      "role": "assistant",
      "content": "<think>...</think>\n<answer>{...}</answer>"
    }
  ]
}

The assistant response contains exactly one <think> block and one <answer> block. The answer JSON includes diagnosis, confidence, focus areas, differential diagnoses, visual evidence, patient-facing explanation, and explicit limitation tokens.

Validation

The canonical train split passed local validation:

Check	Result
Train rows	22,528
Validation rows	2,539
Valid train rows	22,528
Valid validation rows	2,539
Invalid rows	0
SFT conversion skipped rows	0
Missing v2-kept images in generated superset	0
Duplicate image paths	0

Important Limitations

• This repository stores annotation rows with image paths, not redistributed image bytes.
• The dataset has not yet been used to report final model-training results.
• The dataset is intended for research and non-commercial use only.

Citation

If used, cite this dataset together with the source dermatology datasets and the parent reasoning corpus described in the project documentation.