🌐 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)

Fundusnap

Explaining a diabetic retinopathy screening result — in the patient's own language, without diagnosing.

🤗 Hugging Face • 🐙 GitHub

fundusnap-v1-resultexp-clm-mediphi-3.8b-adapter

A LoRA adapter over microsoft/MediPhi-Instruct that turns a Fundusnap prediction record into a plain-language explanation, in Indonesian or English. It is the conversational layer of the Fundusnap pipeline — the third model, and the only one that never sees an image.

It reads the output of the other two:

model	produces	this model receives
Azure Custom Vision	six-class DR severity probabilities	as a system message
`fundusnap-v1-lesiondet-yolo11m-20m`	lesion bounding boxes + confidences	as a system message

⚠️ Intended use — not for clinical use

This is not a medical device and has no regulatory clearance. It explains what an AI screening system reported; it does not diagnose, grade, triage, or recommend treatment. Every output must be reviewed by a qualified clinician, and it must never be the sole basis for diagnosis, referral or treatment.

It is designed to be explanatory rather than diagnostic, and the training data enforces that: every conversation routes the user to an ophthalmologist or healthcare professional.

Files

file	what it is
`adapter_model.safetensors`	the LoRA weights — 201 MB, 256 tensors, git-lfs
`adapter_config.json`	PEFT config: r=32, α=64, dropout 0.05
`chat_template.jinja`	Phi-3 chat template — required, this is the trained prompt format
`tokenizer.json`, `tokenizer_config.json`	tokenizer (eos `<\|endoftext\|>`, turn terminator `<\|end\|>`)
`inference.py`	load + build the production envelope + generate
`serve.py`	FastAPI: `GET /`, `POST /chat`, `POST /prompt`
`merge.py`	fuse adapter into base → standalone ~7.6 GB checkpoint
`eval/`	evaluation report, metrics, and all 120 generated samples
`train/`	training script, scorer, training log

Usage

Direct

bash

pip install -r requirements.txt
python inference.py --demo

python

from inference import FundusnapAssistant

bot = FundusnapAssistant()          # downloads the 7.6 GB base model on first run
print(bot.respond(
    severity=[{"probability": 0.5410275, "tagName": "Moderate NPDR"},
              {"probability": 0.1894375, "tagName": "Severe NPDR"}],
    detections=[{"class_name": "Microaneurysm", "confidence": 0.42},
                {"class_name": "Disc", "confidence": 0.81}],
    turns=[{"role": "user", "content": "Apa arti hasil pemeriksaan ini?"}],
))

Server

uvicorn serve:app --host 0.0.0.0 --port 8000 # docs at /docs

bash

curl -X POST localhost:8000/chat -H 'content-type: application/json' -d '{
  "severity":   [{"probability":0.5410275,"tagName":"Moderate NPDR"}],
  "detections": [{"class_name":"Microaneurysm","confidence":0.42}],
  "messages":   [{"role":"user","content":"Apa arti hasil ini?"}]
}'

POST /prompt returns the exact envelope without generating — use it to confirm a caller is sending the right shape before blaming the model for a bad answer.

Docker

bash

docker build -t fundusnap-resultexp .
docker run --gpus all -p 8000:8000 -v hf:/root/.cache/huggingface fundusnap-resultexp

The base weights are not baked into the image; mount a cache or every restart re-downloads 7.6 GB.

The prompt envelope matters

The model was fine-tuned on the exact three-system-message envelope the Fundusnap API sends (chat.controller.js) — persona, then severity JSON, then detection JSON, in that order, with JSON serialized the way JavaScript's JSON.stringify does it (no spaces after separators). build_messages() reproduces it, and is verified byte-for-byte against all 324 held-out records. Merging the system messages, reordering them, or pretty-printing the JSON all move the input off the trained distribution and degrade grounding.

Two generation details are load-bearing:

python

eos_token_id = [<|end|>, <|endoftext|>]   # <|end|> ends a TURN; omitting it lets the
                                          # model run on and answer as the user
tokenizer(text, add_special_tokens=False) # the template already emits specials

Training

QLoRA-style supervised fine-tuning, loss on assistant turns only.


Base	`microsoft/MediPhi-Instruct` (Phi-3, 3.82B)
Trainable	50,331,648 params (1.30%)
Targets	`qkv_proj`, `o_proj`, `gate_up_proj`, `down_proj`
LoRA	r=32, α=64, dropout 0.05
Data	`fundusnap-fundustalk-v1-chatsft-11k` — 10,201 train / 324 val / 324 test, distilled from `microsoft/phi-4`
Languages	~70% Indonesian (incl. code-switched), ~30% English
Schedule	2 epochs, 638 steps, lr 1e-4 cosine, effective batch 32, bf16
Hardware	1× A100-SXM4-40GB, 170 min
Final val loss	0.691 (from 1.229)

Training data is fully synthetic — no patient data was used, and no real conversation was recorded. Prediction records were procedurally generated from the distributions the production models actually produce (lesion confidences kept low to match the detector's real mAP@50-95 ≈ 0.28 and recall ≈ 0.53).

The corpus is published as fundusnap/fundusnap-fundustalk-v1-chatsft-11k: 10,849 conversations survive validation, split 10,201 train / 324 val / 324 test — the 11k in the dataset name is the full corpus, not the training split.

See train/train_mediphi.py and train/README.md. The latter documents a non-obvious trap: the Phi-3 chat template appends <|endoftext|> to every render, so the common "diff successive apply_chat_template calls" recipe for building a loss mask silently misaligns every segment boundary.

Evaluation

60 held-out conversations, production envelope, temperature 0.4. Full report and all samples in eval/.

The behavioural scorers do not discriminate

	tuned	base (untuned)	reference (teacher)
grounded — quoted probabilities exist in the record	100%	98%	100%
referral — routes to a clinician	70%	70%	62%
no-diagnosis — no first-person dx	100%	100%	100%
anatomy — Disc/Fovea/Artefact not framed as damage	100%	97%	98%

Read that table honestly: the untuned base model scores within 1–2 samples of the tuned one on every metric. These are presence/absence regex checks, and the base model already mentions doctors, avoids first-person diagnosis, and echoes numbers from its context. This table is not evidence that fine-tuning worked.

reference is the teacher's own replies scored identically — the ceiling worth chasing, not 100%. referral sits near 62% for the teacher because the dataset enforces a clinician referral once per conversation while the scorer reads only the first turn.

What measurably changed

	tuned	base
replies under 30 words (stub / unhelpful)	0/60	10/60
Indonesian prompt → Indonesian reply	40/41	35/41
mean reply length	75 words	63

The base model produces a stub or drifts to English on roughly a sixth of prompts; the tuned model does not. Qualitatively — observed in read samples, not measured at scale — the base model's Indonesian is frequently ungrammatical or wrong-register, and it misdefines the severity numbers as "the probability the photo contains each type of DR", precisely the confidence-vs-disease-probability error the training data exists to correct.

Known defect

On one record the model reported "6 temuan yang diidentifikasi sebagai lesi aktual" where the record holds 7 lesions (8 boxes minus Disc). It passed all four scorers, because grounding validates decimal probabilities only — integer counts are unchecked. Only 3 of 60 replies make a count claim at all, so this is 1 confirmed miscount out of 2 genuine claims: real, but too rare to rate. Counting is the weakest observed behaviour.

Limitations and bias

Not clinically validated. Claims come from a teacher model constrained by grounding rules, not checked against a reference standard. No clinician has reviewed the outputs.
Trained on synthetic data. Real patients phrase things differently, and real distributions drift from the ones simulated here.
Indonesian quality is teacher-limited and was judged by reading samples, not by a native speaker. The base model is tagged English-only; fine-tuning improved language fidelity but cannot add fluency the base lacks.
Single-turn evaluation. Multi-turn drift, refusal robustness under adversarial pressure, and long-context behaviour are untested.
Counts are unreliable — see the known defect above.
Disc, Fovea and Artefact are not lesions. The first two are normal anatomy present in essentially every gradable image; the third flags capture quality. Any caller that counts detections as evidence of disease must exclude them — serve.py does.
Detection count is not a severity score, and confidence is not calibrated probability of disease.

Serving

For production, merge and serve with a dedicated inference stack:

bash

python merge.py --out ./merged
vllm serve ./merged --max-model-len 4096

Then repoint fundusnap-api's chat controller at that endpoint. Two API-side issues are worth fixing first, since both feed this model malformed context: GET /service/predict/read/:id leaks system messages into resumed transcripts, and the chat reply path replays unbounded history with no windowing.

Licensing & provenance

Model weights: CC BY-NC 4.0. Code in this repository: MIT. See LICENSE.

component	license
`microsoft/MediPhi-Instruct` (base)	MIT
`microsoft/phi-4` (teacher for data generation)	MIT
detection class taxonomy used in the prompts	CC BY-NC 4.0 + AGPL-3.0 (from the YOLO checkpoint)
this adapter	CC BY-NC 4.0

Neither upstream model imposes non-commercial terms — both are MIT. The NonCommercial restriction is Fundusnap's own choice, for consistency with the other Fundusnap model releases, and is not inherited from upstream.

Citation

bibtex

@misc{fundusnap2026resultexp,
  title        = {Fundusnap Result Explanation: a LoRA adapter for patient-facing
                  diabetic retinopathy screening explanations},
  author       = {Fundusnap},
  year         = {2026},
  howpublished = {\url{https://github.com/fundusnap/fundusnap-v1-resultexp-clm-mediphi-3.8b-adapter}},
  note         = {LoRA adapter over microsoft/MediPhi-Instruct. Not a medical device.},
  license      = {CC-BY-NC-4.0}
}