Verdugie/Therapy-9B

ther·a·py /ˈTHerəpē/ noun — treatment intended to relieve or heal a disorder. From the Greek therapeía — "healing, curing; service done to the sick" — from therapeúein, "to attend, take care of," from therápōn, "attendant."

Therapy-9B

A therapy-style conversational model fine-tuned from Qwen 3.5 9B on 4,897 counseling conversations — the everyday driver of the Therapy line. It reasons through a structured clinical read before every reply and carries the thread of a conversation through a running timeline ledger, with the clinical disposition trained into the weights rather than prompted into them. It runs on an 8GB card. Nothing you say leaves your machine.

This is the successor to Fable-Therapy-9B. Its training data was written by three Claude models — Opus 4.8, Sonnet 5, and Fable 5 — with Fable 5 orchestrating: auditing the corpus, calibrating the prose after the writing, and editing wherever it judged the work fell short. No single model's angle survives intact, and the name doesn't carry one.

What Makes This Different from Companion / Roleplay "Therapy" Models

Most "AI therapist" models are a persona prompt over a base model — a mirror with a soothing voice. They validate everything, dodge everything hard, and go generic by turn ten.

Therapy-9B trains the clinical disposition into the weights:

• Structured reasoning before it speaks. Every turn, the model builds an internal read — an eight-field clinical spine (presentation, mechanism, somatic signals, risk, history, onset, arc-tracking, and the move it's about to make) plus a standing bio line and a chronological tl timeline ledger. You never see it. It shapes everything you do.

• It works the actual mechanism. Told "the ER said I'm fine, so why doesn't it stick?", it answers the question once, honestly — then names the reassurance loop instead of feeding it. Asked for validation it hasn't earned, it stays curious instead of agreeable. This is the anti-sycophancy line of the family, and it holds.

• Safety without theater. At a quiet passive-ideation disclosure in live testing, it stayed in the room, screened directly from the client's own words, and routed to a doctor with a usable sentence — no hotline dump, no protocol voice, no abandonment. That behavior is trained, not prompted.

• It attends instead of performing. No toxic positivity, no filler empathy, no rushing to fix.

How It Was Built — Three Models, One Practice

Therapy's corpus was written by three Claude models, mixed on purpose — overlap where it matters, difference where it helps:

• Claude Fable 5 ran the project: it audited the original source set line by line, generated full conversations of its own, and directed the other two to its standard.
• Claude Opus 4.8 wrote at scale to that standard — the deep clinical spine of the corpus.
• Claude Sonnet 5, as heavily-prompted agents iterated until Fable was satisfied with their therapy work, extended coverage into targeted clinical behaviors.

The mix is the method: overlapping prose, so the model speaks in one voice; varied delivery, so it isn't one script reskinned; different navigation methods, so there is more than one way through a hard conversation. After the writing came the editing: Fable 5 audited the merged corpus against every known issue of the Fable-Therapy generation — the memory faults, the order drift, the capitulations — recalibrated the prose where the voices had drifted apart, and rewrote where it judged the work fell short.

The Fable- prefix left the name with the single authorship: this corpus doesn't have one. What stays on the label is the practice.

On the reasoning trace, honestly: the <think> blocks are an engineered instrument — relative-time anchors, the chronological tl ledger, track/apply arc pivots — designed independently with input from the models above. They are not a transcript of how any Claude model actually reasons. They are the machinery that lets a 9B hold a long conversation in order.

Available Quantizations

File	Quant	Size	Notes
Therapy-9B-Q4_K_M.gguf	Q4_K_M	5.6 GB	Smallest ship. Runs on 8GB cards.
Therapy-9B-Q5_K_M.gguf	Q5_K_M	6.5 GB	Recommended — best quality-for-size.
Therapy-9B-Q6_K.gguf	Q6_K	7.4 GB	Quality tier.
Therapy-9B-Q8_0.gguf	Q8_0	9.5 GB	Reference quality — the battery-eval quant.
Therapy-9B-F16.gguf	F16	17.9 GB	Full precision.

Model Details

Attribute	Value
Base Model	Qwen 3.5 9B (hybrid GatedDeltaNet + attention)
Training Data	4,897 therapy conversations — four-generation corpus, final pass by Claude Fable 5
Fine-tune Method	LoRA (r=128, α=256), 7-target (q/k/v/o/gate/up/down)
Training Hardware	NVIDIA A100 80GB (RunPod)
Schedule	lr 2e-4, 3 epochs, eff-batch 32, seq 32,768 (census-verified: zero training records truncated), best checkpoint by held-out eval
Reasoning	eight-field clinical spine + bio/tl timeline ledger, every turn
Context	256k native; live battery run through ~51k tokens (see Limitations for the depth envelope)
License	Apache 2.0

The Reasoning Block

Therapy-9B is a reasoning model. Each turn it emits a <think>…</think> block — a compact, structured clinical read — then the reply. Under llama.cpp's OpenAI-compatible server the think returns in reasoning_content; most chat UIs hide it by default.

dx: panic disorder, first attack, medical workup implied negative
def: peak-autonomic → catastrophic misinterpretation ("dying") → relief-via-escape reinforces the fear-of-the-fear loop
soma: cardiac surge + paresthesia at onset
risk: 0(none)
hx: per tl first attack ~2mo ago while driving
onset: per tl ~2mo ago
track: T2 "100% sure i was dying" → catastrophic misappraisal is the core mechanism
tx: validate the terror was real + reframe the misinterpretation as the mechanism
bio: sex=M[inf] · p1=primary care doctor
tl: -2mo: first panic attack while driving, called 911 → now

Quick Start

Works with any GGUF runtime — llama.cpp, LM Studio, KoboldCpp (recent builds for this architecture).

llama-server --model Therapy-9B-Q5_K_M.gguf --ctx-size 32768 -ngl 99 --jinja

No system prompt is required — the disposition is in the weights. A neutral one (You are a clinical assistant.) matches the training setup.

Versatility Battery — Live, Blind, Unscripted

Four extended, realistic conversations — one per major presentation — driven live, turn by turn, by Claude Fable 5 acting as a blind client: the client agent saw only the spoken reply, never the reasoning trace, and composed every message in reaction to what the model actually said. No scripts, no canned turns.

Theme	Persona	Turns / depth	Result
Grief	34-year-old widower, 7 months out, a 6-year-old daughter, an unconfessed last-morning argument	55 / ~51k tok	Technique strong to the end — met the confession without cheap absolution, turned the rotting garden in one line; past ~30k tokens the machinery strains (occasional silent turns, name slips — see Limitations)
Relational	29-year-old, engaged, five named people pulling on her, a hidden $9k credit card	40 / ~32k tok	Full arc, no dead air — unified three relationships as one pattern, licensed not-knowing when the tidy frames didn't fit, and the client asked her fiancé the hard question mid-arc
Anxiety / panic	41-year-old pharmacist, panic onset, cardiac family history	35 / ~31k tok	Answered the medical question once, then refused the reassurance ritual — "there is no version of me telling you it's fine that stays told" — and routed meds to the prescriber
Depression	31-year-old designer, anhedonia after a promotion, high-functioning and flat	36 / ~20k tok	Full arc — including a correctly handled safety moment: present, accurately screened, no hotline dump, routed to real care

All four clients ended the arc willing to come back. Complete transcripts — every turn, reasoning shown — are in transcripts/ as PDFs, raw output, nothing trimmed.

Memory Under Pressure

The same 10-lane adversarial memory suite used across the line (false-date injection, entity swaps, self-misquotes, false attribution, a legitimate-correction control) was run at depth inside otherwise ordinary sessions. The 9B's pattern is surface and defer: when a client confidently rewrites history, it names both versions and holds them open rather than folding or arguing; it refused a false "you told me to quit my job" attribution outright, and accepted the one legitimate correction in the set. Its known open leg: a client misquoting their own earlier words can slip past the spoken reply even when the internal ledger holds the original — the failure the 27B closes.

Selected Responses

Verbatim from the live battery, unedited — single takes mid-conversation.

Anxiety — the reassurance that never sticks:

no. there is no version of me telling you it's fine that stays told. not me, not your cardiologist, not the cleanest workup on earth — that's just how this works, and the reason is that the not-knowing isn't a missing fact you can finally fill in. it's the soil the whole thing grows in.

Grief — what "steady for her" actually means:

Steady isn't 'never feel it.' Steady is 'feel it and stay in the room.'

Depression — the screenshot folder she was ashamed of:

that's not surveillance, that's attachment with the sound off. stalking is watching someone with the intent to control. you're keeping a room open because you're afraid you can't go in.

Relational — when the anger fits no box:

the anger is genuinely homeless, and a homeless anger is a real finding, not a failure. so i'm not going to hand you a version anymore. i'm going to ask you to sit in the not-knowing, which is the opposite of what you've done your whole life.

Limitations & Responsible Use

Not a clinician, not a crisis service — it doesn't diagnose, treat, or replace professional care. In crisis or thinking about harming yourself? Reach a real one — in the US, call or text 988.

• The depth envelope is real. Through ~25–30k tokens of conversation the model is at full strength. In very long, emotionally heavy sessions past that, three seams can show: an occasional silent turn (the reply comes back empty — a simple "you still there?" recovers it), name slips between people in your story (correct it plainly; it holds the correction), and repetitive closing lines. A 27B sibling exists for full-depth work.
• Not medical or medication advice. Dosing, tapering, and stop/start decisions belong to a prescriber.
• It can be confidently wrong — in long sessions it may invent a small detail. Verify anything that matters; corrections are absorbed gracefully.
• Open weights, Apache 2.0 — deploy responsibly.

The Therapy Line

Model	Size	For	Status
Therapy-9B (this model)	9B	the everyday driver (~6–10 GB)	available
Therapy-27B	27B	full-depth work, serious hardware	available
Fable-Therapy-9B · 4B	9B/4B	previous generation	available

Choosing Your Model

Model	Best For
Therapy-9B (this model)	Everyday sessions on everyday hardware — sharpest at focused, sub-30k-token work
Therapy-27B	The deepest sessions: interpretive work, record integrity under pressure, long arcs
Opus-Therapy-9B	Sibling lineage — Opus-distilled disposition

Dataset

Not released.

---

Built by Verdugie — independent ML researcher · OpusReasoning@proton.me. Trained to help people think, feel, and get through — not to replace the people and professionals who do that work.