ISEPDistillDataset

Private research repository for the multimodal training corpus used in the ISEP thesis on dermatology-specialized small multimodal language models.

Status: corrected additive source release isep_distill_dataset_v0.4.1 and the derived trainer-ready release e3_multitask_sft_v1 are materialized and validated. The E3 release adds a new configuration without changing the frozen diagnosis, morphology, or caption shards. v0.4.0 remains withdrawn because its task-local split check missed 125 cross-task train/dev group conflicts.

Released configurations

Configuration	sft_train	sft_dev	Total	Target
diagnosis	6,312	1,229	7,541	One canonical label from the frozen 21-class taxonomy
morphology	3,068	527	3,595	All positive concepts from the 48-concept SKINCON ontology
caption	2,767	483	3,250	Short observation-only prefix filtered from authorized SkinCAP text
e3_multitask_sft_v1	25,084	—	25,084	Task-isolated diagnosis, morphology, caption, grounded differential, or image-replacement conversation

diagnosis reproduces the frozen E1 split exactly. morphology excludes 271 SKINCON rows that overlap ISEPDermaBench Validation/Internal groups. Its 606 rows already present in E1 inherit their frozen split; remaining groups use a deterministic 85/15 split with seed 42.

caption excludes 439 author-rejected cases and 243 frozen benchmark overlaps before transformation. A versioned high-precision transform accepts 3,250 of 3,318 technical candidates. The source captions were produced with knowledge of the diagnosis, so these rows are explicitly labelled human_caption_gold_conditioned_filtered; they are not presented as answer-blind descriptions. Written derivative permission was attested by the project owner on 15 August 2026 and the permission document itself is not stored in this repository.

Every caption group now inherits the already frozen E1 or morphology split. The combined diagnosis + morphology + caption release has zero train/dev overlap by leakage_group_id; the manifest records the source of each caption split. This cross-task invariant is checked by the runtime before training.

e3_multitask_sft_v1 expands the 6,312 frozen sft_train diagnosis images into one independently prompted conversation per task. It contains 6,312 diagnosis, 6,312 structured morphology, 6,312 answer-blind caption, 6,127 grounded differential, and 21 request-new-image rows. Stage B accepted 6,148 images and clinically rejected 164; rejected Stage B reasoning is not a trainer-visible target. No distilled sft_dev is published, preserving the frozen development protocol for comparable checkpoint selection.

Objective

The dataset will support two controlled thesis phases after the existing label-only baseline:

• E2_structured: supervision for visual concepts, short captions, grounded differentials, evidence links, uncertainty, and next action.
• E3_hard_kd: the best E2 condition plus filtered response targets generated by a frozen open-weight teacher.

The teacher protocol is two-stage with a strict information boundary. Stage A is answer-blind and describes only what is observable in the image. Stage B receives the same image, the frozen Stage A output, and the private gold diagnosis as a supervised anchor; it produces an auditable differential and a concise student-facing clinical justification without inventing findings.

Configuration roadmap

Configuration	Row unit	Student target	Primary source
diagnosis released	One diagnostic task per image	Canonical label	Normalized gold label
morphology released	One perception task per image	Visible SKINCON concepts	All eligible SKINCON
caption released	One description task per image	Short visible observation	Authorized, filtered SkinCAP prefix
e3_multitask_sft_v1 released	One task-isolated conversation per image/task	Diagnosis, morphology JSON, caption, grounded differential, or request for a better image	Human gold plus accepted frozen Stage A/B targets
structured	One complete clinical task per image	Observations, differential, evidence, uncertainty, action	Accepted Stage A and Stage B
open_response	One open clinical task per image	Short natural-language response consistent with the canonical JSON	Rendering of the accepted structured target

The optional preferences, feature-distillation, logit-distillation, and on-policy artifacts are not part of the first core dataset release.

Dataset Viewer

The release uses explicit dataset-card configs and image-aware Parquet metadata. The Viewer offers diagnosis, morphology, caption, and e3_multitask_sft_v1, and renders the embedded clinical image plus the exact trainer-visible prompt/target messages. Only the three frozen source configs include sft_dev; the derived E3 config contains sft_train only.

config	sample_id	image	target_text	quality_status	messages
diagnosis	source-stable ID	embedded source image	canonical label	accepted	2
morphology	SKINCON-stable ID	embedded source image	deterministic concept JSON	accepted	2
caption	SkinCAP-stable ID	embedded source image	filtered visible observation	accepted	2
e3_multitask_sft_v1	source-stable task-row ID	embedded source image	task-specific accepted target	accepted	2

The synthetic examples under examples/ document future schemas only and are excluded by the explicit configs manifest above.

Common public columns

Group	Fields
Identity	sample_id, case_id, task_id, image_asset_id, view_type
Leakage control	leakage_group_id, split, split_inherited_from_e1, split_source where applicable
Provenance	source_dataset, source_sample_id, license_id, image_sha256
Gold	disease_id, gold_diagnosis, gold_provenance, taxonomy_version
Human morphology	skincon.ontology_version, skincon.source_subset, skincon.positive_concepts, skincon.all_concepts_annotated
Target	target_variant, target_source, config-specific structured fields, messages
Prompt/target	prompt, prompt_sha256, target_text, messages
Quality	schema_version, quality_status

Uneven target coverage

Not every source image will appear in every configuration. Missing supervision is represented by the absence of a task row, never by an invented caption or an all-null placeholder target.

• SKINCON provides 48 binary visual concepts, not prose descriptions. Positive concepts are stored in skincon.positive_concepts; when skincon.all_concepts_annotated=true, every omitted ontology concept is an annotated negative rather than a missing value.
• The reproduced source audit contains 3,866 usable annotations with upstream diagnostic labels: 3,230 Fitzpatrick17k and 636 DDI.
• The training policy admits 3,595 morphology candidates after removing 271 Fitzpatrick17k cases that overlap the frozen internal benchmark. The 480 rows marked Do not consider this image by SKINCON remain excluded.
• In the released morphology config, 1,198 rows map through the versioned taxonomy mapping to the 21 classes. The remaining 2,397 retain valid upstream disease labels but are morphology-only examples; they do not expand the diagnostic output vocabulary silently.
• DDI is now admitted to morphology training. Consequently, DDI results after this release are in-domain/contaminated diagnostics and cannot be presented as independent external generalization evidence.
• caption contains 3,250 filtered SkinCAP observations. Raw captions, diagnoses, and removed suffixes are excluded from trainer-visible shards.
• structured requires accepted Stage A and Stage B components.
• open_response is created only when an accepted canonical structured target can be rendered consistently.
• e3_multitask_sft_v1 keeps diagnosis, morphology, and caption for all 6,312 training images. Only the fourth clinical-response row depends on Stage B acceptance; 164 clinically rejected Stage B samples therefore contribute no grounded-differential or request-new-image target.

The release manifest reports exact shards, hashes, coverage and split counts in metadata/release.json for v0.3 and releases/isep_distill_dataset_v0.4.1/release.json for v0.4.1; the matching quality summaries record the validation gates.

The frozen ontology order and the reproduced coverage audit are stored in metadata/skincon_ontology.json and metadata/skincon_coverage.json. A documentation-only morphology record is available at examples/morphology.example.json.

Only an explicit per-task allowlist determines what is visible to the student. Audit fields must never leak gold diagnosis or private metadata into an answer-blind teacher call.

Private audit material

The following do not belong in the distributable core configs:

• raw private reasoning or chain-of-thought from the teacher;
• long unfiltered teacher outputs;
• free-text reviewer notes or reviewer identity;
• potentially re-identifiable source identifiers;
• sealed benchmark responses and references;
• full-vocabulary logits, hidden states, or feature tensors.

When needed for an experiment, model-side KD artifacts are versioned separately and linked through sample_id, task_id, and generation_id.

Physical layout

ISEPDistillDataset/
  README.md
  data/
    diagnosis/
    morphology/
    caption/                 # withdrawn v0.4.0, retained for auditability
    caption_v0_4_1/          # current Viewer/training shards
    e3_multitask_sft_v1/     # trainer-ready E3 task rows, sft_train only
  examples/
  metadata/
  releases/
    isep_distill_dataset_v0.4.0/
    isep_distill_dataset_v0.4.1/
    e3_multitask_sft_v1/
  schemas/

The Parquet shards were written by versioned builders after split, license, leakage, schema, image-decode, and checksum gates passed. diagnosis, morphology, caption, and e3_multitask_sft_v1 are materialized; structured and open_response remain optional planned configs rather than prerequisites for E3 training.

Versioning policy

• Schema versions follow semantic versioning.
• Every generated release freezes source revisions, taxonomy, prompts, teacher revision, generation parameters, split assignments, and file hashes.
• sample_id + task_id + target_variant is unique inside a release.
• Every image belonging to the same patient, lesion, case, or equivalent image group remains in one leakage_group_id.
• A release tag is created only after automatic audit and manual approval.

License and intended use

This repository is private and research-only. Each row retains its source-level license_id; mixed-source data are not assigned a blanket permissive license. DDI image bytes are included under the user's accepted research-use terms and must not be redistributed publicly. This corpus is not a medical device and is not intended for clinical diagnosis or patient care.