CoolFace
Datasetpublic

TonicAI/synthetic_clinical_conversations

Synthetic Clinical Conversations Fully synthetic English clinical conversations (care-coordination calls, telehealth visits, post-discharge check-ins) paired with structured encounter records, built for training and evaluating transcript→JSON extraction models. Generated structure-first with Tonic Fabricate: the structured facts are authored as relational data with controlled vocabularies, the conversation is rendered from those facts, and the extraction target is a… See the full description on the dataset page: https://huggingface.co/datasets/TonicAI/synthetic_clinical_conversations.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
0likes61downloads
Dataset Card

Synthetic Clinical Conversations

Fully synthetic English clinical conversations (care-coordination calls, telehealth visits, post-discharge check-ins) paired with structured encounter records, built for training and evaluating transcript→JSON extraction models.

Generated structure-first with Tonic Fabricate: the structured facts are authored as relational data with controlled vocabularies, the conversation is rendered from those facts, and the extraction target is a deterministic projection of them — no model ever writes a label. Every extracted fact carries an evidence quote that is an exact substring of a transcript utterance, machine-verified.

Splits

splitrecordsscenariosnotes
train900450clean + ASR-noisy variant per scenario
validation200100rendered by a different model family than train
challenge1005025 counterfactual minimal-pair families (one clinical fact flipped per pair)

Record fields

  • —transcript / utterances — speaker-labeled dialogue (roles: patient, nurse, physician, advancedpracticeclinician, carecoordinator, familycaregiver); variant is clean or production_like (seeded ASR-style noise).
  • —target — the extraction record: encountertype, chiefconcerns, conditions (assertion/temporality/experiencer), symptoms (assertion/severity/onset/duration/ trend/location), medications (dose/unit/route/frequency/status/action/adherence), measurements, followup (incl. `committed`), escalationflags. Controlled vocabularies throughout; negation, family history, corrections, and distractors are designed difficulty features.
  • —evidence — per-fact {fact_id, utterance_id, quote} on clean-variant records, each quote guaranteed to be an exact substring of the referenced utterance (noisy variants share the scenario's facts; their evidence anchors the clean rendering, so the field is empty there).
  • —distractor_names / omitted_names (JSON strings) — facts intentionally excluded from the target: third-party chatter that must not be extracted, and scenario facts the dialogue never states. Useful for precision/trap metrics.

Provenance & privacy

All patients, clinicians, conversations, and identifiers are fictional and machine-generated. No real patient data, transcripts, or PHI were used anywhere in the pipeline. Realism and diversity were enforced with automated gates (template- skeleton deduplication, split-parity audits, evidence verification).

Built for the Tonic.ai cookbook: Zero-label fine-tuning with synthetic conversations (TonicAI/cookbooks).