README.md

metadata

language:
  - en
license: other
tags:
  - rare-disease
  - synthetic
  - clinical-notes
  - phenotype
  - HPO
  - gene-prioritization
  - differential-diagnosis
  - parquet
pretty_name: Rare Disease Synthetic Diagnosis v2

Rare Disease Synthetic Diagnosis (v2)

Fully synthetic rare-disease clinical notes and multi-task diagnosis / gene / variant training rows. Not for clinical use. Contains no real patient data and no RareArena case reports.

Hub note: Assets live on the model repo iLOVE2D/rare_sync (repo_type=model). Hugging Face native load_dataset("{repo}") expects a dataset repo, so it will not resolve this URL automatically. Use download-based loading below (or mirror the parquet trees into a dataset repo if you need Hub configs).

Contents

Config	Split	Rows	Description
cases	train / validation	62,004 / 1,960	Case-level records (HPO phenotypes, genes, variant tables, clinical notes)
tasks	train / validation	132,619 / 600	Multi-task verl-format rows (synth_dx, synth_gene, synth_ddx10, synth_gene10, synth_mcq, synth_variant)
cot_sft	train	10,141	Rejection-sampled chain-of-thought (Qwen3-8B teacher; see caveats)

Note sources among cases: {"llm_rewrite": 63956, "template": 8}.

Load (download-based; model repo)

from pathlib import Path
from huggingface_hub import snapshot_download
from datasets import load_dataset

root = Path(snapshot_download("iLOVE2D/rare_sync", repo_type="model"))
cases = load_dataset("parquet", data_files={
    "train": str(root / "cases" / "train-*.parquet"),
    "validation": str(root / "cases" / "validation-*.parquet"),
})
tasks = load_dataset("parquet", data_files={
    "train": str(root / "tasks" / "train-*.parquet"),
    "validation": str(root / "tasks" / "validation-*.parquet"),
})
cot = load_dataset("parquet", data_files=str(root / "cot_sft" / "train-*.parquet"))

Or read shards directly with pandas.read_parquet / pyarrow.parquet.

Provenance (summary)

• Generator: hpoa_sampler_v2 (pipeline/configs/v2.yaml)
• KB: HPO + HPOA + genes_to_phenotype (see kb_dir in config); HGNC + MONDO refs under data/raw/ref
• Note rewrite teacher: Qwen/Qwen3-8B (6 styles); template fallback when rewrite missing
• Eval holdouts (phenopacket / mygene2 / rarebench / RDS-RDC genes & diseases) excluded at world build
• Val = 3.0% of MONDO groups held out entirely
• Git commit: 862c51d768af48b5032c14005884df5163a8038e
• Packaged: 2026-10-22T00:31:33.654005+00:00
• Full machine-readable provenance: provenance.json in this repo
• Pipeline snapshot: pipeline/ (source + relative configs + DEPENDENCIES.md)

Calibration (difficulty)

Baseline phenotype→disease hit@1 (from data/reports/calibration.jsonl):

Set	hit@1	hit@10	median rank
phenopacket_test_500	0.228	0.464	13
mygene2	0.148	0.345	29
v2_records (synth)	0.543	0.843	1

Synthetic cases are easier than real eval sets; do not treat synth accuracy as clinical performance.

CoT caveat

cot_sft was produced with K=4 rejection sampling + answer-hinted rationalization on unsolved prompts (~76% rationalized). The Qwen3-8B teacher introduces knowledge errors; regenerate with a stronger teacher before relying on CoT for supervised fine-tuning.

What is intentionally excluded

• RareArena / real case-report text (*_mixed exports, sft_ra*)
• Model checkpoints, wandb secrets, API keys
• Raw redistributable KB dumps (rebuild from public HPO/MONDO/HGNC sources)
• Logs, caches, __pycache__

License / use

Ontology sources retain their upstream licenses (HPO, MONDO, HGNC, OMIM identifiers as references). Synthetic text is provided for research and model training only. Do not use for diagnosis or patient care.

Reproduction

# after cloning pipeline/ or the parent project
python scripts/generate_v2.py
sbatch scripts/vllm_rewrite.sbatch
python scripts/export_v2.py
python scripts/package_hf_dataset.py --upload --repo_id iLOVE2D/rare_sync