lookevink/scribeski-intake-dialogues

Scribeski Intake Dialogues

Clinician–client conversations labelled against a 63-field behavioral-health intake form: demographics, housing and household, work and income, substances, medications, PHQ-9 and GAD-7 items, suicide/violence risk items, daily-living items, referrals and follow-up. It was built to train a small classifier that reads a whole session transcript and answers every form field at once, leaving a field blank unless the conversation actually establishes it.

No client data is included. The default intake config is fully synthetic behavioral-health sessions, and is the only part used to train Scribeski's classifier.

Content note: some synthetic sessions include discussion of suicidal thoughts, self-harm, substance use and domestic violence, written to resemble clinical risk assessments.

What's inside

subset	source	language	conversations
synthetic	synth-v1	en	90
synthetic	synth-v1	es	22
synthetic	synth-v2	en	889
synthetic	synth-v2	es	201

Train / validation: intake/train 998, intake/validation 204. Total transcript length ≈ 52,035 minutes.

subset	what it is	labels come from
intake	Behavioral-health intake, follow-up and crisis sessions (English and Spanish) written by open-weight LLMs from a sampled ground-truth sheet	the sheet (by rule), kept only where a blind teacher model reproduces them

Train/validation is split by conversation group: a restyled variant always sits in the same split as its original. Use the group column if you make your own splits.

How it was made

Synthetic sessions (tools/synth/generate.mjs):

1. Code samples a ground-truth sheet. For every field it draws a value and how it's disclosed: stated, indirect, corrected later, misreflected by the worker, only hypothetical, about someone else, a past state, pending (applied for / an earlier referral), ambiguous, declined, or never discussed. The label follows from the disclosure mode by rule; for example, hypothetical, someone else's, and never-discussed all mean blank.
2. An LLM invents a persona consistent with the sheet.
3. An LLM writes the dialogue ~7 minutes at a time from a brief that says what to establish, how, and what must never come up. Sessions vary in presenting problem, setting, client and worker style, structure, speaking rate (70–170 wpm) and transcript surface (clean or ASR-like).
4. A separate blind pass extracts every field. agreed is true only where it matches the sheet.

Writers: DeepSeek V4 Flash, gpt-oss-120b, Qwen3-Next-80B-A3B, Inkling-Small, Hy4-preview, and (early batches) Gemma 4 31B, all open-weight under permissive licenses. Blind teacher: DeepSeek V4 Flash. Batches marked synth-v1 predate the variety features and used DeepSeek only.

Fields

schema/fields.json lists all 63 fields with their options and plain-language meaning. Each record's labels is a list of:

key	meaning
field	form field id
value	list of option codes; [] = not established (leave blank); multi-select fields may have several
agreed	train only on agreed: true; false means the label sources disagreed
mode	how the fact was disclosed (synthetic), or REAL / RESTYLED
acceptable	other values that also count as correct ("" = blank), e.g. NOT_ASSESSED for an unasked risk item

transcript is a list of {time, speaker, text} turns. provenance_json holds generation details (writer, teacher, sheet summary, scenario, origin of a variant).

Limitations

• Synthetic text is LLM-written. It is less messy than real speech even with the style controls, and writers sometimes drift into topics the brief forbade. Those labels are filtered out (agreed: false), but the drift stays in the text.
• Not validated by clinicians. Not for clinical use.

Authors

Kevin Loo, Columbia University · kevin@loo.ski

License and attribution