danielfdias98/derm-reasoning-next-stage-v2

Dermatology Next-Stage Reasoning v2

This dataset contains JSON-only chat SFT targets for a dermatology vision-language router. It is part of a dissertation project on privacy-preserving multi-agent dermatology support systems.

The model role is deliberately narrow: it inspects a dermatology image, extracts visible evidence, checks image quality, and routes the case to a broad category-level specialist agent. It is not trained here to provide final medical advice.

Dataset Tracks

Config	Rows	Train	Validation	Purpose
routing_v2	22,501	20,251	2,250	Main visible-evidence category-routing dataset.
context_ver_v2	4,019	3,617	402	Single-turn uncertainty dataset that asks for missing context before routing.
context_dialogue_v2	2,500	2,250	250	Two-turn clarification dataset with disclosed synthetic context.
cot_ablation_v2	10	9	1	Small controlled ablation for explicit reasoning-trace format.
swi_reasoning_v2	1,000	900	100	Controlled ablation for adaptive minimal/explicit reasoning depth.

Total accepted rows: 30,030.

Category Routes

The router uses broad category agents:

• inflammatory
• benign
• malignant_or_premalignant
• infectious
• undetermined

Rows must not route to disease-specific agents.

Format

Each row is a chat SFT object with messages and metadata. The assistant content is a JSON string.

The production routing schema includes:

• visible_evidence
• image_quality
• decision
• reasoning_summary
• differential
• needs_clarification
• clarifying_questions
• red_flags_to_check
• limitations
• user_message

The swi_reasoning_v2 ablation adds:

• reasoning_mode
• switch_reason

The user_message is a routing handoff, not a confirmed diagnosis or treatment recommendation.

Images

This repository contains annotations and chat targets only. It does not embed the underlying dermatology images. Image references in messages[*].content point to the project image layout, for example final/train/....

Use the source image repositories or the dissertation project's local image workspace when training a VLM that requires image files.

Splits

Image-disjoint train/validation splits are stored in splits/. The split manifest is available at splits/split_manifest.json.

Validation

Rows were generated in small Codex sub-agent batches and accepted only after schema validation, SFT conversion, and text linting.

Key local checks:

• accepted export JSONL files parse successfully;
• train/validation splits have zero source_image overlap per track;
• context_dialogue_v2 text lint: 2,500 rows, zero findings;
• context_dialogue_v2 structural check: 2,500 rows, zero malformed dialogue rows;
• swi_reasoning_v2 schema validation: 1,000 rows, zero validation errors;
• swi_reasoning_v2 text lint: 1,000 rows, zero findings;
• final routing audit report: audits/final_routing_v2_spot_audit.md.

Intended Use

Recommended experimental order:

Base VLM
-> Label-only SFT checkpoint
-> VER Routing v2 continuation checkpoint
-> Context-Seeking VER v2 continuation checkpoint
-> Context-Seeking Dialogue v2 continuation checkpoint

cot_ablation_v2 and swi_reasoning_v2 should remain separate unless evaluation shows that they improve accuracy and safety without increasing hallucination, verbosity, unsafe certainty, or invalid JSON.

Limitations

• This dataset is annotation-generated and requires downstream model evaluation.
• It does not replace clinical validation, clinician review, dermoscopy, pathology, or prospective testing.
• Context-Seeking rows ask for missing clinical context but do not contain real patient history unless explicitly provided in a disclosed synthetic dialogue example.
• The dataset is for routing and structured evidence extraction, not autonomous diagnosis or treatment.