m0rtyddd/clario-synth-v3

Clario synthetic patient diaries, v3

Short, colloquial patient diary entries in English and Italian. Each entry is labelled with the symptoms it mentions, mapped to the Human Phenotype Ontology (HPO). For every symptom the label gives the exact quote, whether it is negated ("no fever"), whether it belongs to a relative and not the patient ("my mother has..."), and onset and severity hints where the text gives them.

It was built to fine-tune the patient-side extractor of Clario, a tool that turns everyday notes ("my knuckles are stiff when I wake up") into clinical terms with a quoted source. No public dataset combines lay language, span-level symptoms, negation and HPO ids (see docs/DATA_SOURCES.md in the repo), so this one was generated.

Every row is synthetic. No real patient, record or medical text was used. Labels are correct by construction (see below), not by human annotation. This is not a clinical resource and must not be used to diagnose anyone.

Contents

Split	Rows	Source diseases	English / Italian	Hard negatives	Labelled entities
train	2,392	2,308	1,346 / 1,046	433	9,917
validation	300	133	168 / 132	46	1,272
test	300	150	168 / 132	49	1,237

A hard negative is an ordinary diary entry with no health complaint at all; its target is an empty entity list.

The split is held out by disease: no source disease appears in more than one split, so the test set measures diseases the model never saw. It is also stratified by language. 8 duplicate diaries were dropped out of 3,000 generated rows (SPLIT.md).

Files:

File	What it is
train.jsonl, val.jsonl, test.jsonl	One row per diary: id, diary, target, source_disease, lang, kind (clinical or hard_negative), generator
*_sft.jsonl	The same rows as chat messages (system prompt, diary, expected JSON answer) ready for supervised fine-tuning
lay_lexicon.json	Lay lexicon: 5,071 everyday English phrasings for 726 HPO terms, {hpo_id: [phrases]}
lay_lexicon.json.stats.json, synth.jsonl.stats.json, run.log	Generation statistics and the full run log
SPLIT.md	Split counts written by prep.py

target.entities[] fields: quote (verbatim span from the diary), canonical (HPO label), hpo_id, type, negated, family_history, onset_hint, severity (0 to 10 or null), confidence (always null here).

Example (Italian, truncated):

{"diary": "Lun 12 Giu. Oggi ho finalmente ridipinto il corridoio, ma sono distrutta. ... Poi c'è questo dolore alla pancia 6/10 ...",
 "target": {"entities": [{"quote": "dolore alla pancia", "canonical": "Abdominal pain", "hpo_id": "HP:0002027",
   "type": "symptom", "negated": false, "family_history": false, "onset_hint": null, "severity": 6, "confidence": null}, "..."]},
 "source_disease": "ORPHA:424016", "lang": "it", "kind": "clinical", "generator": "Gemma-4-31B-IT"}

How it was made

All generation used Gemma 4 31B IT (Gemma-4-31B-IT, open weights, Apache 2.0) through the AMD Token Factory API, on AMD GPUs, on 16 and 17 October 2026. Every row records its generator in the generator field. No Claude, GPT or other closed-model output is in the data: the scripts refuse Anthropic endpoints.

The pipeline lives in train/ of the Clario repo:

1. Lay lexicon (lay_expand.py). HPO terms have few lay synonyms, and those sound clinical. For each of 800 frequent terms Gemma wrote about 10 phrasings in different voices (teenager, pensioner, with typos, tied to a body part). Each phrase had to pass two checks: a code check (no medical vocabulary, not a copy of the term) and a blind round trip, where the model sees the phrase next to 5 similar HPO terms and has to pick the original. 7,700 phrases passed the code check, 4,917 passed the round trip (64%), and terms with no surviving phrase fall back to HPO synonyms.
2. Diary entries (synth.py). For each row the script picks a rare disease from HPO annotations (1,822 OMIM and 1,178 Orphanet diseases across the 3,000 rows), some of its symptoms, which ones are present, negated or belong to a relative, optional onset and severity cues, unrelated life events as distractors, and a writing voice. Gemma writes the entry. Code then checks that every symptom is quoted verbatim, negations and relatives are written correctly, required cues are present and no clinical label leaked in. Rows that fail are thrown away, so the labels come from the sampling plan, not from the model.
3. Split (prep.py), as described above.

Generation numbers: 3,000 rows accepted out of about 6,550 attempts (46% acceptance). Main reject reasons: a symptom not mentioned (1,278), the relative/patient attribution wrong (692), a quote not verbatim (473), a negation wrong (381), a clinical term used instead of lay words (298). About 3.0M input and 1.9M output tokens for the diaries, roughly 8.3 hours at the API's shared rate limit. The lexicon took about 1.9 hours more.

To reproduce, follow docs/ENRICHMENT_RU.md in the repo:

cd train
python lay_expand.py --base-url $TF_BASE_URL --api-key-env TF_API_KEY --model Gemma-4-31B-IT --terms 800
python synth.py      --base-url $TF_BASE_URL --api-key-env TF_API_KEY --model Gemma-4-31B-IT --n 3000
python prep.py

Sampling is random (temperature 1.0), so a rerun gives different text with the same label scheme.

Training

The dataset is meant for a LoRA fine-tune of Gemma 4 E4B as Clario's on-device extractor (train/train_rocm.py, optional GRPO in train/train_grpo.py). At the time of upload no model has been trained on it yet, so there are no results to report here. Evaluation numbers, when they exist, will be in eval/results/ of the repo, never copied by hand.

from datasets import load_dataset
ds  = load_dataset("m0rtyddd/clario-synth-v3")          # diary + structured target
sft = load_dataset("m0rtyddd/clario-synth-v3", "sft")   # chat messages for SFT

Known limitations

• Synthetic text. It reads like a diary but is more regular than real patient writing, and the model sometimes slips in semi-clinical wording ("the scar tissue in my liver") that the filters did not catch.
• Symptom combinations follow disease annotations, so they are rarer and denser than everyday complaints.
• The lay lexicon is English only, so Italian entries rely on ontology wording plus the model's own phrasing.
• Labels are only as good as the checks. They are not reviewed by a clinician. Measure real accuracy on human-reviewed text (Clario's gold review page is eval/gold/review.html; the reviewed set is not part of this upload).

Licences and credits

This card uses license: other because the rows combine sources:

• Text was generated by Gemma 4 31B IT (Apache 2.0 model licence).
• Labels (HPO ids and labels) come from the Human Phenotype Ontology. HPO's licence asks for citation and version, and forbids altering the HPO files; no HPO file is redistributed here. Cite: Gargano et al., The Human Phenotype Ontology in 2024, Nucleic Acids Research. https://hpo.jax.org
• source_disease ids are Orphanet (ORPHA:) and OMIM (OMIM:) identifiers reached through the HPO annotation file phenotype.hpoa. Orphanet is credited as a source of rare disease data (https://www.orpha.net). OMIM is a trademark of Johns Hopkins University (https://omim.org); no OMIM text is included, only ids.

Generated with free AMD Token Factory credits during a hackathon. The Token Factory terms describe the service as trial and development only.