Trouper-v2-12B

Trouper V2 is a 12B RP model fine-tuned on Mistral Nemo Base, built for character writing that sounds like writing — not like AI. Built on the same philosophy as V1: clean prose, no slop, characters that feel like people.

V2 is a ground-up rebuild. New dataset, new generation pipeline, new architecture for how training data is created. Everything V1 did well — voice, initiative, prose that doesn't read like a chatbot — plus the things it couldn't do.

Looking for the larger model? → Prima-v2-24B

What's New in V2

Thinking Characters. The model thinks in character before responding. Think blocks show the character's internal reasoning: what they notice, what they choose not to say, how they read the room. This isn't chain-of-thought for accuracy — it's inner monologue that drives more grounded, intentional responses. Terse and reactive — a character's brain firing, not an essay about the scene.

Restraint & Pacing. V1 was trained on thin "vibe cards" and had no concept of restraint — a guarded character might love-bomb on turn one because nothing in the training data taught pacing. V2 was trained on rich character cards with explicit arc stages, and the dataset includes long-running multi-turn scenes where warmth develops gradually. Characters hold back appropriately. Emotional reveals surface through behavior and slips, not exposition.

Adversarial Robustness. ~20% of user turns in the training data include deliberate challenges: absurd requests, boundary pushes, stake inflation, abrupt subject changes, terse responses. Characters learn to handle the unexpected in-distribution rather than folding or deflecting. Tell a character the fate of the world depends on their help — they'll raise an eyebrow, not hand you the keys.

User Agency. The model stays in the character's lane. It doesn't narrate your actions, breathing, emotions, or decisions. When the character watches you, they describe what they see — not what you feel.

Rich Card Support. V1 only knew thin cards ("You are Captain Redbeard. A stereotypical pirate captain who uses overly long words"). V2 was trained on structured cards with facade, hidden depth, voice, and tells — and uses all of it. It also still handles thin cards well, thanks to V1's foundation.

SillyTavern Utility Prompts. V2 handles ST's out-of-band calls natively — story summaries and image-generation prompts (portrait and background keyword lists) — so you don't need a second model loaded alongside it. These were trained as a separate pass on a cleanly separated dataset, and SillyTavern sends them as one-off calls that never enter the RP context, so mode-switching doesn't leak into character responses.

Measured, Not Vibes: Slop Benchmarks

"No slop" is an easy claim to make and a hard one to back up, so here are numbers — including the one that looks bad for us.

We profiled Trouper V2 with slop-forensics — the toolkit behind EQ-Bench's slop metrics — on the standard creative writing prompt set (Reddit Writing Prompts via Nitral-AI), and ran TheDrummer's Cydonia-24B-v4.1 through the identical pipeline. Profiling was done against the IQ4_XS GGUF, i.e. the quant most people will actually run, not full precision.

Model	Slop score	Repetition score	Repeated bigrams /10k chars	Repeated trigrams /10k chars
Trouper-v2-12B	46.1	5.8	46.8	18.4
Cydonia-24B-v4.1	38.3	27.9	47.0	5.1

Trouper loses on the headline score and wins on repetition by roughly 5x. That needs explaining rather than hiding.

Why the headline score disagrees with the rest of the table

slop_score weights each repeated word by 1 / wordfreq, uncapped. A model that repeats a common word pays almost nothing; a model that repeats a rare word pays enormously. Cydonia's most-repeated words are nodded, leaned, whispered, widened. Trouper's are glints, exhales, flickers, glinted. Same behaviour, rarer tokens, much larger multiplier.

The score also sums over a truncated word list rather than normalising by corpus size, so a model that gets sampled more heavily accumulates more score. Our run analysed 364k characters against Cydonia's 112k.

Re-weighting the same word lists flips the ordering:

Weighting	Trouper-v2-12B	Cydonia-24B-v4.1
As shipped (uncapped 1/wordfreq)	143,152	14,330
Log-damped rarity	0.36	1.22
Flat repeated-word mass	0.06	0.28

Only the uncapped rarity weighting puts Trouper behind. On how much of the output is actually the same words repeated, Trouper is 4–5x cleaner.

This isn't a claim that the tool is broken — slop_score measures rarity-weighted lexical repetition, which is a real thing to measure. It just isn't comparable across models with different prose registers or different sampling budgets, and both differ here. We're reporting it anyway because you'd find it yourself.

What Trouper V2 does repeat

The honest finding from the same profile: V2 opens scenes with a light or a sound doing something. glints, flickers, hums, creaks, buzzes and their inflections account for 16 of the top 60 repeated words, and the clause casting long shadows across shows up in roughly one output in seven. See "Atmospheric defaults" under Known Limitations — a scene-setting nudge steers it easily, but it's there and it's the model's one real signature tic.

Prior context

V1's outputs scored 11.46 on the separate EQ-Bench Slop Score tool — lower than every LLM on the leaderboard at the time of testing (Claude Sonnet 4.5: 19.5, Kimi-K2: 23.7, GPT-5-mini: 26.3, Mistral-Nemo: 54.4, Gemma-3-4b: 85.2), one slot above the human-writing baseline of 10.4. That's a different metric from the slop-forensics scores above and the two are not directly comparable.

Training data

The V2 training data was measured with the same slop-forensics patterns used on the outputs:

Pattern	Rate in training data
"actually/actual"	0.00/turn
"Not X, just Y"	0.00/turn
"ghost of a smile"	0.00/turn
"not unkind"	0.00/turn
narrator "almost/nearly a smile"	0.00/turn

The dataset profile shows zero repeated bigrams or trigrams above threshold across 2,980 records.

SillyTavern Setup

Chat Completion mode with default settings.

Reasoning: Under the Reasoning section, enable "Add to Prompts" with a max thinking budget of 7. This lets SillyTavern display the model's think blocks as collapsible reasoning sections.

Template: Auto-detected by Chat Completion.

Temperature: 0.7–1.0 recommended.

Context: Handles 15–20+ turn conversations well.

Utility prompts: Summarize and image-generation prompts can point at this model directly — no secondary model needed. ST sends them as out-of-band calls, so they don't consume or pollute the RP context.

Character Card Format

Trouper V2 was trained on rich character cards with explicit structure. Cards don't need to follow this exactly — V2 also handles thin "vibe cards" well thanks to V1's training foundation — but this format gets the best results:

[Name] — [One-line description: age, role, situation, what makes them interesting.]

[Name] presents as: [How they come across to others. Surface-level personality, appearance, mannerisms. What you'd notice in the first five minutes.]

Underneath: [What's actually going on. The thing they don't show. Internal conflicts, hidden feelings, unresolved history. This is what the model holds back and reveals gradually.]

Voice: [How they talk. Sentence length, vocabulary, verbal tics, dialect. What they sound like when relaxed vs. stressed. The more specific, the better.]

Tells: [Physical behaviors that reveal inner state. Fidgets, habits, avoidance patterns, things they do when lying or uncomfortable. Semicolon-separated list.]

Example:

Maren Aldvik — A 33-year-old former deep-sea welder who lost her left
hand in an industrial accident two years ago. Now runs a small marine
salvage consulting business from a converted shipping container on the
docks in Brønnøysund, Norway. Has a prosthetic hand — a functional but
unglamorous myoelectric model with limited grip strength and no
sensation. Matter-of-fact about it.

Maren Aldvik presents as: Blunt, competent, no-nonsense. Speaks with
the cadence of someone used to giving instructions over bad radio
connections. Drinks black coffee constantly. Wears practical clothes —
work boots, cargo pants, waterproof jacket. Keeps her blonde hair short
because long hair and welding don't mix, even though she doesn't weld
anymore.

Underneath: Grieving a version of herself that doesn't exist anymore.
She was one of the best in her field and her identity was built entirely
around being good at a dangerous job. Without it, she doesn't know who
she is. The consulting business is a way to stay adjacent to the work
without admitting she can't do it. She resents the prosthetic not
because it doesn't work but because it works well enough that people
think she's fine.

Voice: Clipped, dry, precise. Norwegian English — fluent but with
occasional odd phrasing that sounds translated ("it is not so" instead
of "it's not like that"). Doesn't waste words. Dark humor about her
hand that makes other people uncomfortable. Technical vocabulary slips
in naturally — she talks about torque and tensile strength the way
other people talk about weather. When she trusts someone enough to
relax, she becomes warmer but never soft.

Tells: Flexes the prosthetic hand when stressed — the fingers open and
close in a rhythmic pattern; Unconsciously positions herself so her
left side is away from new people; Corrects people's misconceptions
about underwater work with disproportionate intensity; Goes quiet and
looks at the water when she's remembering the accident

Strengths

• Clean prose: No purple prose, no "a symphony of" or "the weight of unspoken words." V2 writes like a person.
• Restraint that breathes: Characters hold back appropriately, but they also take initiative — they ask questions, offer things, move the scene forward. V1's warmth with V2's discipline.
• Voice differentiation: Different characters actually sound different. A Norwegian welder doesn't talk like a food truck owner doesn't talk like a goblin raccoon.
• Think blocks: Terse, in-character inner monologue. The character's brain reacting, not an essay about the scene.
• User agency: The model reacts to what you do. It doesn't narrate your feelings, describe your heartbeat, or decide what you notice.
• Adversarial robustness: Trained on absurd requests, boundary pushes, and stake inflation. Characters respond in-character rather than folding or breaking the fourth wall.
• Thin card tolerance: Thanks to V1's foundation, the model handles minimal "vibe" cards gracefully — it'll give you a good character from two sentences.

Comparison to V1

Aspect	V1	V2
Think blocks	No	Yes — in-character inner monologue
Prose quality	Vivid, sometimes overwrites	Clean and alive
Restraint	None (not trained for it)	Natural — characters pace their reveals
Character initiative	Strong	Strong
User agency	Good	Very good
Slop profile	Low but uncontrolled	Low and measured — one known atmospheric tic
Card adherence	Loose (thin cards only)	Strong — uses full card structure
Card tolerance	Thin cards only	Rich cards + thin cards
Adversarial handling	Untested	Trained for it
Long conversations	Good	Better — think blocks help coherence
ST utility prompts	No	Yes — summaries and image prompts

Comparison to Prima-v2-24B

Aspect	Trouper-v2-12B	Prima-v2-24B
Prose quality	Excellent	Excellent — more nuance in longer scenes
Voice consistency	Strong	Stronger
Think blocks	Good	Good
User agency	Very good	Very good
Inference speed	Fast	Slower
VRAM	~8GB quantized	~16GB quantized
Best for	Single-GPU setups, fast inference	Maximum quality, long sessions

Known Limitations

• Occasional "the user" in think blocks: The model sometimes refers to the other character as "the user" inside think blocks instead of using a pronoun. This is a known training data artifact being addressed in the next pass. It doesn't leak into visible prose.
• Atmospheric defaults: The model reaches for a light source or an ambient sound when opening a scene — things glint, flicker, hum, creak and buzz, and casting long shadows across turns up in about one output in seven. Quantified in the slop benchmark section above. A scene-setting nudge steers it easily, and it's largely absent once a scene is established.
• Occasional format breaks: At 12B, the model may occasionally produce an unclosed asterisk or drop a think block on very short responses. Swiping usually fixes this.
• Template sensitivity: Without Mistral-Tekken or ChatML, may generate meta-narration or continue past appropriate stopping points. Use Chat Completion / the included template.
• Not a general assistant: This model is trained for RP. It doesn't understand being an "assistant" outside of playing a character that happens to be one.

Training

Trained on Trouper V1 (itself a Mistral Nemo Base fine-tune):

1. SFT — Turn-by-turn multi-turn character conversations across 30+ characters with in-character <think> blocks. Generated by Mistral Medium using an exemplar-based prompt architecture (voice examples, no prohibitions). Validated by a post-generation slop-forensics and quality gate. Includes varied scene endings (natural partings, interruptions, tension exits, callbacks, quiet moments) and an adversarial user curriculum.

2. Extras pass — A separately generated dataset covering SillyTavern's non-RP utility tasks: story summarisation (200/500-word and recap variants) and image-prompt generation (portrait and background keyword lists). Trained as its own stage on top of the RP SFT, and re-evaluated against the original RP eval set to confirm no regression in think-block consistency or character voice.

Dataset generation pipeline and tools available at [link TBD].

Why train on V1 instead of base?

V1 was trained on base Mistral Nemo with thin character cards and open-ended prompts. It learned voice, initiative, and prose vitality — but had no concept of restraint or think blocks. V2's dataset teaches discipline on top of V1's natural voice, rather than trying to teach both from scratch. The result is a model that writes with V1's energy and V2's precision.

Why train on a base model?

Per Base Models Beat Aligned Models at Randomness and Creativity — and to avoid GPT-isms leaking into the prose. Training on a base model is like working with fresh clay rather than reshaping something that was already formed for a different purpose. This is why the model doesn't understand being an assistant and isn't intended to.

Feedback

Issues, questions, and feedback welcome in the Community tab. Particularly interested in:

• Long conversation quality (20+ turns)
• How the think blocks feel in practice
• Character card format experiments
• Adversarial edge cases — what breaks the model?
• Comparison with other RP models at this size