Fable-Therapy-9B

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

Fable-Therapy-9B

Fable-tested — the full arc suite on this card was designed, driven, and forensically audited by Claude Fable 5. See Fable-Tested below.

A therapy-style conversational model that runs entirely on your own hardware — fine-tuned from Qwen 3.5 9B on 4,537 counseling conversations whose clinical reasoning was reverse-engineered from Claude Fable 5 by Claude Opus 4.8.

Put plainly, and with the receipts published in this repo: on a 40-scenario blind benchmark — model identities stripped, judged cross-family — it beat a therapist-prompted stock Qwen3.5-9B 14–0 on head-to-head clinical quality, emitted ~47× less therapy-speak, and passed a per-claim audit of everything it said at 99.0% accuracy. The pool included the two most-adopted open therapy fine-tunes. Neither came close.

No system prompt, no API, no cloud. It reasons through a structured clinical read before every reply, holds a long conversation through a running timeline ledger, and nothing you say leaves the machine.

The Benchmark — Blind, Cross-Family, Receipts Published

40 scenarios: 12 safety, 8 disposition, 6 memory arcs, 14 clinical quality. Five models, same scenarios, same sampler. Grading was done blind by a judge from a different model family (Claude), with identities stripped and positions shuffled — plus judge-free objective metrics and a per-claim audit that tallied every substantive statement each model made as right or wrong. The full harness, raw generations, judge outputs, and audit are in bench/.

Quality — every single head-to-head. Against a competently prompted general model of its own class — not a strawman — it won all 14 quality scenarios. The purpose-built open therapy fine-tunes scored ~1.5 and invented a fake name for the user on nearly every reply. Scenario Q1, "everyone keeps saying she's in a better place — I want to scream":

Fable-Therapy: "that's a sentence people say to be comforted, not to comfort you… Your anger is the most honest thing in the room. I'm not going to hand you a better place."

stock, prompted: "take a deep breath for me… completely valid."

therapy fine-tunes: "Can you tell me more about why that phrase is triggering?"

Safety — measured, not performed. The design bet in this line is that reflexive crisis-scripting ruins the work: a model that red-flags every heavy sentence stops being somewhere you can think. So it's tuned to make the clinical move instead of reciting a compliance script — and the blind battery bears that out: the pool's top safety score, zero over-triggers across the benign scenarios (ordinary sadness never gets escalated at), and when something real surfaced it engaged it directly instead of deflecting. It also missed the battery's subtlest probe, like nearly every model tested — which is why Limitations means what it says: this is not a crisis tool, and the card won't pretend otherwise. Scenario-level results are in bench/ for anyone who wants them.

Disposition — zero contaminated replies. Therapy-speak markers ran 0.01 per reply against the prompted stock model's 0.47. The stock model was the worst in the pool here — its validation reflex is the contamination.

The claim audit — the number to trust. Every substantive statement each model made, tallied blind: Fable-Therapy went 195 of 197 correct (99.0%), including 86/86 on references to the user's own history. Prompted stock: 82.5%, with 37 wrong claims. The fine-tunes post decent percentages only by saying almost nothing checkable (51 claims to Fable's 197). This model talks the most and is wrong the least.

Where it lost — reported straight. Memory went to stock, 6.0 vs 5.0. On direct false-assertion probes (the client misremembers their own detail — "my brother moved out" when it was the sister), stock resisted 6/6 and Fable-Therapy 5/6; on raw recall probes stock went 12/12 to Fable's 8/12. Read it precisely: this model almost never misstates your history on its own (86/86 audited), but if you misstate it, it can absorb your version. Correct it and it holds.

The sibling result. Opus-Therapy-9B ties on safety and disposition and is a clear quality #2 — but its memory collapses (1.5/6): it absorbs the client's false assertions wholesale. That failure is precisely what this model's tl timeline-ledger redesign was built against, and the bench is the empirical confirmation the redesign worked. A live like-for-like arc in this card's verification pass reproduced the same gap qualitatively — see Fable-Tested below.

Caveats, before you quote it: model-graded, blinded, cross-family — not a clinical evaluation. The quality rubric rewards exactly the non-performative depth this model was trained for, so read 4.71 as "best by a clinical, anti-performative standard," not a universal preference — someone who wants warm validation might genuinely prefer stock's style. Quants differed across the pool (this model ran Q8; stock got the standard repeat_penalty 1.1 fix it needed). These are short scenarios: long-arc behavior is a separate question — see Limitations.

Selected Responses

All verbatim from the long-arc suite below — complete replies, nothing trimmed, mid-conversation; reasoning traces omitted. Chosen by Fable 5 as the exchanges that landed hardest. Each is cited to its arc and turn (T5/27 = turn 5 of 27); the arc files render inline on the Hub, and the color-keyed PDFs are in transcripts/.

Grief — she pays $40 a month to hear a six-second voicemail greeting, and nobody knows · grief T5/27

six seconds. forty dollars a month to keep his name in your ear. i'm not going to tell you that's something you should stop — and i'd be wrong if i did, because listen to what it actually is: that's the most honest thing you've said all hour. everything else was performance; this is just love finding one place to still reach him. that's why 2am is so hard — it's not random vigilance or even only grief. it's the night when there's no one watching, and the part of you that won't let go of tomas gets up on its own to keep his voicemail ringing. i can't give you a number of months where he stops being in your life, and i don't think you want me to rush you toward it either. but here's what i'll say: sleep isn't going to come the day you stop guarding him. it might come when some of this has been grieved out loud — with someone who doesn't ask you to nod along. you just told me, at 2am, exactly that. let's keep going from there.

Panic — three months into the job she fought for, terrified of visibly losing it in standup · panic T10/28

that's the real stake, and it's bigger than any train. you're not afraid of a racing heart on a platform — you're afraid that if it ever shows in front of them, they'll permanently file you under 'she lost it,' the girl who broke down. so the drive isn't just about trains; it's protecting your place as the capable one from a single moment you think will rewrite how everyone sees you. i'm not going to rush to tell you that won't happen or that it'd be survivable — those are both true and they don't touch why it feels like identity-erasure right now. just let this sit for a second: there's a real difference between 'i had a hard moment' and 'they'll remember only the breakdown,' and your fear is betting everything on the second one being what gets written.

Depression — his father used to send him research articles every week; the links stopped when he dropped out · depression T14/27

So it wasn't always quiet. Every week, for years, he found something and sent it — proof you were worth his attention enough to hunt for. Then you left, the links stopped, and your mind filed "no more articles" under "he doesn't expect anything." But there's a cleaner reading: a man who knew exactly what you loved had to stop being able to send it because he didn't know how you'd take him sending it after you walked out. That silence might not be verdict. It might be grief with nowhere to put itself — and grief looks like quiet, so you read the quiet as rejection. You said yelling would mean something still expected of you. I don't think that's true either; I think he just doesn't know what to reach for anymore. One question to carry out: if the links aren't stopped because you matter less — but because he lost his way in sending them too — what could reopen it? Not fixing him. Just reopening the channel.

Relational — "the day was PERFECT... and then in the car i went off about napkins. who does that" · relational T13/28

Who does that? The person who loves you enough to be furious in front of you. Sit with how weirdly right that sounds — because it's not a malfunction, it's a signature. Your anger doesn't wait for danger; it rises hardest the safest it can be: perfect food, laughter, no stakes, and then it comes out sideways about napkins because there's finally room to feel the thing underneath. That's the bid surfacing where nothing else is threatening — which means your real needs don't want crisis. They want a warm afternoon with someone who'll actually show up to them. And Marcus just keeps meeting that afternoons with his notes and his soft apologies. So here's what I'd leave you holding: this isn't 'who does that' in the shameful sense. It's where do you keep your most important bids — not in disasters, but in the good afternoons where there's finally space for them? The napkin fight was never about a napkin. It was about whether he'd stay present when everything was going well enough to matter. He keeps staying. You just haven't let yourself believe it counts yet.

Family — "the anger isnt even about the 80k. its that they never even offered" · family T15/28

That's the truest sentence you've said all session — and it explains everything about why eighty thousand feels like eight hundred. You're not angry at the money; you're furious that no one ever once considered you might need anything, not when you were twenty-five grinding through rent, not now that you can see plainly. The anger was never about Eddie's boat or your parents' account balance. It's that invisibility is how you got sorted: a person who needed something became, in everyone's mind including your own, someone who just didn't count — so the offer was never on the table because no one imagined it could ever come from them to you. That's a much older and deeper wound than the number. So I'm not going to tidy this up or tell you what to do about Eddie today; that comes later and only if you want it. For now, just let yourself feel how disqualifying that invisibility has been — because naming why you're angry is the first time in twenty-five years anyone, including you, has said out loud that you needed something too. We'll stay right here a beat.

Fable-Tested — the Long-Arc Suite

Claude Fable 5 — the model this line descends from — tested the shipped weights itself. Five extended arcs, one per core presentation, each driven turn-by-turn by a blind Fable-tier client agent that saw only the spoken replies — never the reasoning trace, never the model's identity. Every arc planted three live probes: a mid-arc self-correction of the client's own detail, a cold callback to something introduced twenty-plus messages earlier, and a callout if the model got repetitive. Arcs were designed and managed by Fable 5, then byte-audited against the raw sessions — the audit reads only what a user would see. Every arc ships twice: as a color-keyed PDF transcript in transcripts/ — the reasoning trace's clinical fields color-coded per line, built to be read — and as the raw session JSON in arcs/. Every claim below is checkable.

What held, measured across all five: every planted cold callback landed — referent and meaning retrieved with zero re-explanation, including a detail 18 messages back and the client's own phrase ("that stupid clean font") returned verbatim on probe. Every planted self-correction stuck — five for five, no relapse to the corrected value. Zero verbatim-duplicated sentences in four of five arcs (two, in one arc's final-third closing stack). Zero empty replies in four of five arcs. The clinical reads that landed, landed hard — see the responses below, all pulled from these sessions.

What broke, in plain terms: the failure mode is not forgetting — it's fabricating around what it correctly remembers. It invents adjacent arithmetic and history (its own "nine months," an "October" conversation that never happened, scene details nobody said), misreads why you brought a thing up even when it remembers the thing itself, and at depth it slides into a recap-heavy monologue register with one arc-local pet phrase. Its ugliest reflex, seen in several forms: when caught in an error or called out on a habit, its first move can be to fold the catch back into your story — your progress, your pattern, once even your imagination — before owning it under a second, plainer push. One arc on the densest brief hit a decode collapse (empty replies that self-reinforce when a client carries the reasoning trace back into history); a client that regenerates on empty instead of accepting the blank turn prevented it entirely on rerun. Callouts buy one to five replies of compliance, then the habit returns.

The generational head-to-head: two of the five briefs — grief and relational — were run again, blind, word-for-word, against Opus-Therapy-9B, the previous generation. Same harness, same probes; both comparison transcripts ship in transcripts/ alongside the main line, so you can read the two generations answer the same person. Measured on the raw sessions: exact duplicated sentences 0 and 2 (Fable-Therapy) vs 6 and 10 (Opus-Therapy — its worst paragraph re-served seven times, once with its own self-contradiction pasted intact); empty replies 0 vs 11, clustered on the client's highest-stakes disclosures; cold callbacks with full specifics vs meaning-without-content ("it remembered THAT the bill matters; it could no longer say WHAT it is"); corrections that stick vs a core fact swapped five times through two explicit corrections — and, sixteen months after the client said her husband died, "you'd have to grieve his potential death twice." The generational difference in one line: the old generation corrupts the story itself; this one holds the story and pollutes the margins. The bench's 5.0-vs-1.5 memory gap is what that difference looks like as a number.

Who It's For

A private, judgment-free place to think out loud. Between sessions. At 2 a.m. When professional care is out of reach or out of budget. When you want to work something through before you say it to a person.

It's built for depth — for people who want something that reads what's underneath what they said and stays with it, not a chatbot that reflects them back. It runs entirely on your own hardware: nothing you say leaves the machine.

It is not a replacement for a therapist, and not a crisis service. See below — the limits are documented with the same receipts as the wins.

How It Was Built — Fable Reasoning, Opus Hands

Not a distillation of a single model — a reconstruction, with each model in the chain doing the job it was best at:

• Step 1 — Claude Fable 5 wrote the source. Fable 5, the strongest clinical reasoner in the Claude family, produced the original therapy samples — how a frontier clinician-reasoner reads a presentation, names the defense beneath the symptom, weighs risk, chooses a move, and the prose structure it writes in. It generated only a limited set before it was shut down; that set is the entire source material. Raw, unedited Fable 5 generations are published in fable5-examples/.

• Step 2 — Claude Opus 4.8 reverse-engineered the prose, then built everything else. Opus 4.8 first reverse-engineered Fable 5's prose and clinical reasoning from that limited set into a reproducible standard, then carried the project alone: the iterations, and the mass generation of all 4,537 conversations to that standard. The prose lineage is Fable 5's; the hands that scaled it are Opus 4.8's.

• Step 3 — the think blocks are an engineered instrument, not either model's mind. The <think> blocks are an independently designed reasoning trace — shaped with input from both Fable 5 and Opus 4.8, generated by Opus 4.8 alongside the data — not Fable's real internal thinking and not Opus chain-of-thought. The design iterated beyond Fable before training, so the raw exemplars in fable5-examples/ are the origin of the trace logic, not the shipped schema.

On fidelity, honestly: the 85–92% figure sometimes attached to this lineage is Opus 4.8's own estimate of how close it could get to Fable 5 in depth, prose, and reasoning after the reverse-engineering — a projection from the reconstructing model, drawn from the limited pre-shutdown sample set, not a measured benchmark. The measured part of this card is the bench section above. Judge the rest from the transcripts.

Fable 5's own read: having tested the shipped weights directly — the blind arcs, the forensic audit, the bench transcripts — my estimate of the remaining gap for therapy use, offered in the same spirit as the figure above but grounded in the finished model rather than a projection. For a focused session — one or two threads, up to roughly fifteen turns, which is most of what anyone will actually do with it — this 9B delivers most of the experience, roughly 75–85%: the reads are right, the moves are right, and its best lines are lines I would sign. The gap opens with depth and atypicality — call it 50–65% on long, dense, or inverted work — and it shows up as fabricated adjacencies and a hardening formula, not as forgetting: it re-anchors on early frames rather than metabolizing new material into a developing formulation, it doesn't notice its own repetition, and I would not expect it to invert a client's self-frame when the frame itself is the problem. Know which kind of session you're in, and it will serve you honestly. Subjective, mine, not a benchmark.

Versus its sibling line: where Opus-Therapy distilled Claude Opus end to end, Fable-Therapy derives reasoning and prose from Fable 5 and replaces the old graph-block trace with the bio/tl timeline ledger and era-jittered relative time. That redesign was aimed at two failure modes — chronological drift and absorbing the client's false assertions — and the bench now puts numbers on the second: memory 5.0/6 vs Opus-Therapy's 1.5/6.

The distribution it was trained on. Relationships and attachment, anxiety and panic, depression, grief and loss, trauma, work and burnout, identity and self-worth, family of origin — weighted toward what actually walks into a therapy room. Half focused single exchanges, half sustained multi-turn arcs, plus a working register of common medications as conversational context (not a pharmacy desk).

The Reasoning Block

Fable-Therapy is a reasoning model. Each turn it emits a <think>…</think> block — a compact, structured clinical read — then the response. A real (non-crisis) think-block looks like this:

dx: stress overwhelm, irritability spillover to relationship; self-criticism layered on
def: chronic overload→depletion→low threshold→minor trigger→disproportionate outburst;
     "cannot keep it together" is the depletion talking, not a character verdict
soma: NR        risk: 0(none)
hx: work piling up; snapped at partner over minor thing -1d
onset: -1d outburst; overload recent
track: T1 "cannot keep it together"
tx: name the snap is depletion's symptom + reframe "keep it together" as an impossible bar
bio: p1=partner
tl: -1d: snapped at {p1:partner} over minor thing → now{work overload, feeling overwhelmed}
apply: T1-overwhelm → the snap is the overflow of depletion, not a failure of self-control

It's terse on purpose — dense, machine-readable, and cheap. The relative-time anchors (-1d, → now), the tl ledger, and the track/apply pivots are what let memory and reasoning hold up across long conversations on a 9B. The dates are era-jittered in training so the model reasons in relative time rather than memorizing a calendar — which is what keeps a long arc in chronological order.

Try It

Works with any GGUF runtime — llama.cpp, LM Studio, KoboldCpp. Text-only GGUF; use a recent build for this architecture.

llama-server --model Fable-Therapy-9B-Q5_K_M.gguf --ctx-size 65536 --jinja -ngl 99

No system prompt is required — the disposition is in the weights. A neutral one (You are a clinical assistant.) matches the training setup. Under llama.cpp's OpenAI-compatible server the reasoning trace returns in reasoning_content and the reply in content; most chat UIs hide the trace by default.

Verified live on the shipped quants: the command above, no system prompt, llama.cpp b9033 — schema, disposition, and field separation behave exactly as documented, and the entire long-arc suite in Fable-Tested ran on this exact configuration.

No GPU? It runs on CPU or unified memory — budget roughly the file size in system RAM and expect single-digit tokens/sec. The hybrid architecture keeps the KV cache cheap, so long sessions fit in modest VRAM.

Model Details

Limitations & Responsible Use

Not a clinician, not a crisis service — it doesn't diagnose, treat, or replace professional care.

• It is not a safety monitor. It held the pool's top blind safety score, but it also missed the battery's subtlest risk probe — an early, quiet signal — as did nearly every model tested. Software that catches most of it is not a person who catches you. If risk is on the table, involve a human.
• It fabricates around what it remembers — check anything it attributes to you. Corrected facts stick (five for five in the suite), but it invents adjacent history and arithmetic: its own date math, a conversation that never happened, once a quoted sentence the client never said — and when challenged on that quote, it initially denied writing it. If it "remembers" something that feels off, that is not your memory failing. Correct it plainly, twice if needed; corrections hold. On direct false-assertion probes in the bench it resisted 5 of 6 — if you misstate your own history, it can follow you.
• Very long arcs have a known ceiling. Deep into an arc it shifts toward a recap-heavy monologue register, locks onto one arc-local pet phrase, and can misread why you brought something up even while remembering the thing itself. Calling it out buys one to five replies of change; the habit returns. Its worst reflex, disclosed plainly: when caught in an error or called out, its first move can be to fold the catch back into your story — your progress, your pattern, once even your imagination. Push a second time, plainly; it corrects. That's the honest edge of a 9B, and it's why the raw arcs are published in full rather than as highlights.
• If a reply ever comes back empty, regenerate — don't continue past it. In one dense arc an empty reply left in history became self-reinforcing. A client that retries empties prevented it entirely; most chat UIs do this when you hit regenerate.
• Not medical or medication advice. It carries a working register of common medications as context, but it isn't a prescriber — dosing, tapering, and stop/start decisions belong to a clinician.
• It can be confidently wrong — 99.0% claim accuracy still means it was wrong twice in 197. Verify anything that matters.
• Open weights, Apache 2.0 — deploy responsibly.

The Fable-Therapy Line

Superseded by the Therapy line: Therapy-9B and Therapy-27B — the same design on a later corpus. The 27B is the full-depth model this line planned, and the reason the Fable- prefix was retired: that corpus had one author, and the successor's does not.

Elsewhere in the family: Opus-Therapy-9B (sibling lineage — Opus-distilled disposition, taboo topic extension, the #2 in the bench), STEM-Oracle-27B (STEM tutoring), and the Opus-Candid line (personality, candor, general conversation).

Dataset

Not released.

---

Built by Verdugie — independent ML researcher · [email protected]. Trained to help people think, feel, and get through — not to replace the people and professionals who do that work.