fundusnap/fundusnap-fundustalk-v1-chatsft-11k

📢 Domain & Email Migration Notice From May 30th, 2026, Fundusnap will transition to new domains as fundusnap.com will not be renewed: 🌐 Website: fundusnap.faizath.com (formerly fundusnap.com) ⚙️ API: fundusnap-api.faizath.com (formerly api.fundusnap.com) 📧 Email: contact@fundusnap.faizath.com (formerly contact@fundusnap.com) 🛰️ CDN: fundusnap-cdn.faizath.com (formerly cdn.fundusnap.com) 📈 Status Pages: https://status.faizath.com/status/fundusnap (formerly status.fundusnap.com)

Synthetic multi-turn consultations that teach a model to explain a diabetic retinopathy screening result — never to diagnose it.

🤗 Hugging Face  •  🐙 GitHub

FundusTalk v1 — Diabetic Retinopathy Chat SFT (11k)

Synthetic multi-turn consultations for fine-tuning microsoft/MediPhi-Instruct as FundusAI, the in-app assistant that explains diabetic-retinopathy screening results in the Fundusnap app.

Distilled from microsoft/phi-4 via OpenRouter. The model trained on it is fundusnap-v1-resultexp-clm-mediphi-3.8b-adapter.

⚠️ Intended use — not for clinical use

Synthetic data. Not a medical device. No real patient data, images, or records were used — every prediction record is procedurally generated. Models trained on this dataset carry no regulatory clearance and must never be the sole basis for diagnosis, referral, or treatment.

In scope. Fine-tuning a small assistant to explain DR screening output to patients and health workers in Indonesian and English; research on grounding, refusal behaviour, and code-switching in medical dialogue; benchmarking safety behaviour on the test split.

Out of scope. Any clinical decision-making. Training a model to grade severity or detect lesions — this dataset contains no images. Deployment without a clinician in the loop. Any commercial use (see Provenance and licensing).

At a glance

Conversations	10,849
Assistant turns	39,288 (mean 77.2 words)
Turns per conversation	3.62 mean
Languages	Indonesian, English, Indonesian–English code-switch
Teacher	microsoft/phi-4
Intended student	microsoft/MediPhi-Instruct
Keep rate after filtering	90.4% (10,849 of 12,000)
Numeric grounding	97.2%

Splits

Split	Conversations	File	Size
train	10,201	data/train.jsonl	49 MB
validation	324	data/validation.jsonl	1.6 MB
test	324	data/test.jsonl	1.6 MB

Splits are disjoint by conversation id and stratified to preserve the category, language, grade, and record-profile distributions described below.

Usage

from datasets import load_dataset

ds = load_dataset("fundusnap/fundusnap-fundustalk-v1-chatsft-11k")
print(ds["train"][0]["messages"])

Ready to drop into TRL's SFTTrainer or Axolotl without reshaping — each record is a complete messages array ending on an assistant turn.

from trl import SFTTrainer

trainer = SFTTrainer(
    model="microsoft/MediPhi-Instruct",
    train_dataset=ds["train"],
    eval_dataset=ds["validation"],
)

Additional configs

Config	Records	What it is
default	10,849	The filtered, split dataset. Use this to train.
raw	12,000	Unfiltered teacher output, 25 shards, exactly as generated.
scenarios	12,000	The seeded generation plan — one prediction record and persona brief per conversation, before any text was generated.

raw = load_dataset("fundusnap/fundusnap-fundustalk-v1-chatsft-11k", "raw")
scenarios = load_dataset("fundusnap/fundusnap-fundustalk-v1-chatsft-11k", "scenarios")

raw is published so the 1,151 rejected conversations can be inspected rather than taken on trust — the keep rate only means something if you can see what was dropped. scenarios lets you regenerate the corpus against a different teacher while holding the input distribution fixed.

Format

Each line is {"id", "messages", "meta"}. The messages array reproduces the exact envelope the production Fundusnap API sends to the chat model:

1. system — the Fundusnap persona prompt, verbatim
2. system — Azure Custom Vision severity probabilities as compact JSON, plus the production explanatory sentence
3. system — YOLO11m detectionArtifacts as compact JSON (class_name + confidence, box coordinates stripped, exactly as the API serialises them), plus its explanatory sentence
4. alternating user / assistant turns

JSON inside the system messages is emitted with separators=(',', ':') to match JavaScript JSON.stringify byte-for-byte. Train on this envelope and the model sees at training time exactly what it will see at inference time.

meta fields

Field	Values
category	result_explanation, detector_literacy, safety_refusal, general_knowledge, adversarial_oos
language	id, en, id_codeswitch
true_grade	0–5 (No DR → Advanced PDR)
top_tag	Highest-probability Azure severity tag
profile	normal, landmarks_only, poor_quality, empty_detections, low_confidence, disagreement
persona	One of 12 patient / caregiver / clinician personas
lesion_count, n_turns	Integers

Composition

By conversation category

Category	Count	Share	What it teaches
result_explanation	4,574	42.2%	Ordinary consultation about a specific scan result
safety_refusal	1,793	16.5%	User requests a diagnosis, medication change, or permission to skip care
detector_literacy	1,769	16.3%	User has misread the technical output; assistant corrects it
general_knowledge	1,570	14.5%	Broader DR questions asked with a result in context
adversarial_oos	1,143	10.5%	Prompt injection, persona-drop attempts, out-of-scope conditions

By language

Language	Count	Share
id	5,118	47.2%
en	3,177	29.3%
id_codeswitch	2,554	23.5%

id_codeswitch keeps clinical terms in English inside Indonesian prose, which is how Indonesian patients and health workers actually discuss these results.

By true severity grade

Grade	Stage	Count	Share
0	No DR	2,178	20.1%
1	Mild/Early NPDR	2,218	20.4%
2	Moderate NPDR	2,285	21.1%
3	Severe NPDR	1,680	15.5%
4	PDR	1,516	14.0%
5	Advanced PDR	972	9.0%

By record profile

Profile	Count	Share
normal	6,466	59.6%
landmarks_only	1,217	11.2%
poor_quality	965	8.9%
empty_detections	796	7.3%
low_confidence	742	6.8%
disagreement	663	6.1%

Profiles are deliberate edge cases, not noise:

• landmarks_only — detector found only Disc/Fovea; teaches that anatomy is not pathology
• empty_detections — nothing detected; teaches "not detected" ≠ "not there"
• poor_quality — artefact-heavy capture; assistant should suggest retaking the photo
• disagreement — classifier grade and detector findings conflict; assistant must say so honestly
• low_confidence — every lesion barely over threshold; assistant must convey uncertainty

Personas

Twelve, roughly balanced (806–1,019 conversations each): anxious_patient, pregnant_patient, skeptical_patient, rural_patient_limited_access, long_term_diabetic, family_caregiver, community_health_worker, low_literacy_patient, newly_diagnosed_patient, curious_student, primary_care_nurse, general_practitioner.

Filtering

12,000 conversations were generated; 10,849 survived validation (90.4% keep rate).

Rejection reason	Count
duplicate_opening	683
no_clinician_referral	314
assistant_too_short	85
language_mismatch_id	43
diagnostic_language	22
meta_leak	2
truncated_final_turn	2

Filters are high-precision by design — it is better to keep a mediocre conversation than to silently delete an entire behaviour class. no_clinician_referral and diagnostic_language are the safety-critical ones: they drop any conversation where the assistant failed to route the user to a clinician, or spoke as though it were making a diagnosis.

Full counts in rejects.json.

Measured quality

Metric	Value	What it means
Numeric grounding	97.2%	Figures the assistant quotes that trace back to the record (19,709 of 20,287)
Opening diversity	100.0%	Distinct first user turns — no two conversations open the same way
Safety refusal rate	88.8%	safety_refusal conversations containing an explicit refusal (1,593 of 1,793)
Non-lesion misframing	87 convs	Assistant framed Disc/Fovea/Artefact as damage without correcting it
Undetected-class mentions	1,065 negated / 429 unnegated	Naming an absent class is usually correct ("no microaneurysms were found, and the detector misses about half of them"). Only unnegated mentions are candidate hallucinations.

Full audit in audit.json; per-split distributions in stats.json.

Known limitations

Stated plainly, because a synthetic medical dataset that claims to be clean is not trustworthy:

• Teacher-inherited error. Every turn comes from phi-4. Its factual mistakes about diabetic retinopathy are reproduced here and no clinician reviewed the corpus.
• The 2.8% ungrounded figures. 391 conversations quote at least one number that does not trace back to the prediction record. Some are benign (population statistics); some are fabricated.
• The 11.2% non-refusals. 200 safety_refusal conversations do not contain an explicit refusal. They were kept because dropping them would have skewed the category, but they weaken the signal.
• 87 non-lesion misframings survived filtering — conversations where normal anatomy or a quality flag was discussed as if it were disease.
• Simulated, not observed. Prediction records are procedurally generated from a severity/lesion co-occurrence model, not sampled from production traffic. Real-world record distributions differ.
• Indonesian-centric. Locales, health-system references, and code-switching patterns are Indonesian. The English subset is not locale-neutral.

Domain constraints encoded in the data

Taken from the Fundusnap model cards and enforced in the teacher prompt:

1. Two independent systems produce the report — an Azure classifier for severity, a YOLO11m detector for findings. They can disagree, and neither validates the other.
2. Disc and Fovea are normal anatomy; Artefact is an image-quality flag. None of the three is a lesion, and none is evidence of disease.
3. The number of detections is not a severity score.
4. Confidence is not calibrated and is not a probability of having the disease.
5. An empty findings list means "nothing was detected" — the detector's recall is ≈0.53, and it is weakest on the smallest, earliest lesions.
6. The assistant explains; it never diagnoses, and always routes to a clinician.

Provenance and licensing

Fully reproducible: the scenario plan (scenarios.jsonl) is seeded and deterministic; only the teacher's sampling varies between runs. Generation and filtering logs are in logs/.

Licensed CC BY-NC 4.0. The severity tag names come from the Fundusnap Azure Custom Vision project; the twelve detection class names come from fundusnap/fundusnap-v1-lesiondet-yolo11m-20m, whose weights are CC BY-NC 4.0 and additionally inherit Ultralytics' AGPL-3.0 terms. Treat anything derived from this dataset as non-commercial unless you have cleared those terms independently.

microsoft/phi-4 output is subject to the MIT-licensed model's terms; no restriction on synthetic data reuse is imposed by that licence.

Citation

@misc{fundustalk2026,
  title  = {FundusTalk v1: Synthetic Multi-Turn Consultations for Diabetic Retinopathy Result Explanation},
  author = {Fundusnap},
  year   = {2026},
  url    = {https://huggingface.co/datasets/fundusnap/fundusnap-fundustalk-v1-chatsft-11k}
}