JulianAT/medlyn-synth-v1

Medlyn synthetic clinical corpus v1

Summary

100,000 synthetic patient records, German documents and English fact labels. No real patients or challenge documents are included. This release is prepared privately first.

Config	Train	Validation	Test
full	80000	10000	10000
10k	8001	999	1000
verifier	5001	600	900

Generation pipeline

Seeded structured histories become documentation plans and fresh German clinical templates. Controlled mutations change side, dose, date, assertion, BMI arithmetic or schedules. Labels are derived from final rendered text and mutation ledgers, never Medlyn's extractor. Bounded LLM prose rewrites retain their source split and must preserve planned values and spans. Generator version: 1.0.0; seed: 20261030; rewritten patients: 300.

Schema

full and 10k: one row per patient: id, family, split, documents (list of structs with id, title, date, type, department, text), plus JSON string columns facts, mutations, contradictions, distractors, gaps. Decode JSON columns with json.loads. Document text retains source <BR>/entity formatting. Fact offsets use UTF-16 code units into engine-normalised text, not raw text or Python character offsets. verifier: one claim/passage example per row: patient_id, family, split, source, claim, label, assertion, timing, mutation, document_id, fact_id, quote_hash, start, end, and JSON string columns passages and messages. Labels: supported, unsupported, contradicted. Sources are synthetic only.

Splits and leakage control

Patient and generation-family isolation is checked within each config. The 10k config is a deterministic largest-remainder sample stratified by family and split; it retains the full config's assignments. Subset selection uses SHA-256 ranking with seed 37. Held-out styles, scenarios and mutations follow the source manifests. Verifier input snapshot: datasets. Verifier training has its own pre-augmentation family assignment. Do not concatenate configs for evaluation: full and verifier configs contain 37 shared patients assigned to different splits. Use each config independently or re-split the entire union by family before training. Before writing any text-bearing row, the exporter scans every string against all eight local challenge documents. The scanner uses the check-data normalisation and rejects every contiguous 120-character overlap, including unaligned windows. scan-report.json records counts and reference hashes without source text. Real verifier examples and predictions are never discovered or copied. Release file hashes are in release-manifest.json.

Label tiers

These labels are intended_code_checked: planned values, final spans and controlled mutations pass code checks. They are not independently semantically verified or human gold. Candidate extractor outputs are not truth. The PRD's verified and human-reviewed gold tiers require additional independent review that this release does not claim.

Intended use

Research on evidence-grounded claim verification, retrieval, contradiction handling and controlled evaluation. CC BY 4.0 applies to this synthetic release; attribute Medlyn.

Limitations

Synthetic patients and simplified German style templates do not represent real clinical diversity. English claims and narrow mutation patterns can create shortcuts. Document rewrites do not establish clinical validity. No clinical review or patient-care validation is implied. Do not use this corpus to justify autonomous clinical decisions.

Citation

@misc{medlyn_synth_v1,
  author = {Schmidt, Julian},
  title = {Medlyn synthetic clinical corpus v1},
  year = {2026},
  url = {https://huggingface.co/datasets/JulianAT/medlyn-synth-v1}
}