Mediform/german-consultation-dialogues

Synthetic German Doctor–Patient Consultations

Realistic, synthetic German doctor–patient consultation transcripts, generated locally for the Scribion medical-documentation pipeline (ASR → fusion → fact extraction → note). Each transcript is grounded in real medical Q&A from Mediform/german-medical-sft and deliberately exercises the Scribion ASR boost vocabulary (conditions, symptoms, and German drug brand names).

⚠️ Synthetic data. Generated by QuantTrio/gemma-4-31B-it-AWQ, grounded in a validated medical-QA corpus. No real patients. Useful for scaling, controllable coverage, and privacy; not a substitute for real recordings in final ASR acoustic validation (no real accents / overlap / mis-hearings).

How it was built

1. Seeds — medical QA from Mediform/german-medical-sft tagged with Scribion boost terms; grouped into scenario bundles (a specific-condition anchor + clinically co-occurring comorbidities, via a term co-occurrence graph).
2. Generation — gemma-4-31B-AWQ (+ MTP assistant draft, spec decoding) writes a spoken-style consultation grounded in the bundle facts, with Arzt:/Patient: turns and structure (Begrüßung → Anliegen → Anamnese → Untersuchung → Beurteilung → Plan), weaving in the boost vocabulary incl. drug brand names.
3. Gate — heuristic validity (format, turns, no truncation/meta-leak) + sampled LLM correctness/realism judge.

Schema

• id, anchor (chief topic), patient {anrede, name, age}
• turns: [{speaker: doctor|patient, text}]
• transcript: full text
• boost_terms, n_boost_terms: Scribion boost terms the transcript exercises
• scenario_terms: the clinical scenario (anchor + comorbidities)
• grounding_facts: the source QA facts the consultation is grounded in

Stats

• 4,800 consultations (after heuristic + LLM-judge filtering of 5,079 generated)
• 1,693 distinct clinical scenarios · ~4,400 distinct patients · ages 19–88
• avg 25.5 speaker turns/dialogue
• 771 distinct Scribion boost terms exercised (≈4.9/dialogue, appearing naturally)
• (A larger, more diverse 15k edition is at Mediform/german-consultation-dialogues-15k.)
• Every record carries LLM-judge quality scores (1–5). Set means: realism 4.58 · medical_correctness 4.65 · coherence 4.99 · grounding 4.94
• Dangerous/low-quality dialogues (judge dangerous_error / keep=false / medical<3) were removed (generation dangerous-error rate was 1.0%).

Method note (quality)

An earlier v1 that forced a boost-term list per dialogue scored realism 2.69 and 10.7% dangerous (keyword-stuffing / impossible cases like a female with prostate disease). The shipped v2 dropped the forced term list (vocab appears naturally), made patient sex coherent with the condition, and instructed conservative, guideline-near medicine — raising realism to 4.58 and cutting dangerous to 1.0%. This trades boost-term density (≈5 vs ≈20/dialogue) for realism — the right tradeoff for ASR/documentation transcripts.