Instructions to use arbazsiddiqui/Ozan-v1-12B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- llama-cpp-python
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with llama-cpp-python:
# !pip install llama-cpp-python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="arbazsiddiqui/Ozan-v1-12B-GGUF", filename="ozan-dpo200-Q4_K_M.gguf", )
llm.create_chat_completion( messages = [ { "role": "user", "content": "What is the capital of France?" } ] ) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M # Run inference directly in the terminal: ./llama-cli -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama-cli -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Use Docker
docker model run hf.co/arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
- LM Studio
- Jan
- vLLM
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "arbazsiddiqui/Ozan-v1-12B-GGUF" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "arbazsiddiqui/Ozan-v1-12B-GGUF", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
- Ollama
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with Ollama:
ollama run hf.co/arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
- Unsloth Studio
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with Unsloth Studio:
Install Unsloth Studio (macOS, Linux, WSL)
curl -fsSL https://unsloth.ai/install.sh | sh # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for arbazsiddiqui/Ozan-v1-12B-GGUF to start chatting
Install Unsloth Studio (Windows)
irm https://unsloth.ai/install.ps1 | iex # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for arbazsiddiqui/Ozan-v1-12B-GGUF to start chatting
Using HuggingFace Spaces for Unsloth
# No setup required # Open https://huggingface.co/spaces/unsloth/studio in your browser # Search for arbazsiddiqui/Ozan-v1-12B-GGUF to start chatting
- Pi
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Configure the model in Pi
# Install Pi: npm install -g @mariozechner/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:8080/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Hermes Agent new
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:8080/v1 hermes config set model.default arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Run Hermes
hermes
- Atomic Chat new
- OpenClaw new
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:8080/v1 \ --custom-model-id "arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
- Docker Model Runner
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with Docker Model Runner:
docker model run hf.co/arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
- Lemonade
How to use arbazsiddiqui/Ozan-v1-12B-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull arbazsiddiqui/Ozan-v1-12B-GGUF:Q4_K_M
Run and chat with the model
lemonade run user.Ozan-v1-12B-GGUF-Q4_K_M
List all available models
lemonade list
Ozan-v1-12B
A 12B creative-writing model tuned to write long-form fiction that doesn't sound like an LLM.
Ozan means "storyteller / bard" in Turkish. The model is a fine-tune of Mistral-Nemo-Instruct-2407 built around one goal: low slop. "Slop" is the family of over-used LLM tells: "barely above a whisper," "a testament to," "couldn't help but," the reflexive "not just X, but Y." It comes mostly from training on GPT-style synthetic data, and it's the single biggest reason machine prose reads as machine prose. Ozan is trained, and measured, to avoid it.
- ๐ฆ GGUF quants: arbazsiddiqui/Ozan-v1-12B-GGUF
- Full-precision (BF16): arbazsiddiqui/Ozan-v1-12B
- Code & recipe: github.com/arbazsiddiqui/Ozan
Why this model exists
Instruct models tell a decent story but write sloppily: base Mistral-Nemo-Instruct scores 9.47 on the EQ-Bench Creative Writing v3 slop metric. Two steps fix that:
- SFT (QLoRA) on ~17k curated, low-slop examples: prose distilled from strong non-GPT models (the Claude-Opus and Gemini-3-Pro families) plus filtered human Reddit WritingPrompts for range. This teaches voice and roughly halves the slop.
- Gutenberg anti-slop DPO: preference-tuned toward real public-domain book prose over AI prose. This crushes the slop that's left without hurting fluency.
Results: lowest-slop runnable 12B
Slop & repetition are computed directly from the generated text (no LLM grading involved) by the EQ-Bench Creative Writing v3 harness (github.com/EQ-bench/creative-writing-bench, by Sam Paech), run identically on every model's 96 generations (32 prompts ร 3). Lower is better.
| model | size | slop โ | repetition โ |
|---|---|---|---|
| gemma-2-9b-it | 9B | 11.92 | 6.30 |
| Mistral-Nemo-Instruct (our base) | 12B | 9.47 | 3.93 |
| Magnum-v4-12b | 12B | 8.57 | 8.99 |
| Gemma-3-Glitter-12B | 12B | 7.68 | 3.63 |
| Gemma-3-Starshine-12B | 12B | 7.52 | 3.59 |
| gemma-3-27b-it | 27B | 7.03 | 3.62 |
| Ozan-v1-12B | 12B | 5.30 | 3.21 |
| QwQ-32B | 32B | 4.90 | 3.56 |
| Mistral-Small-Creative-24B | 24B | 4.25 | 3.36 |
Lowest slop of any same-size 12B tested, beating the popular community finetunes and even a 27B, with the cleanest repetition in the whole field and zero degenerate loops across 96 stories.
Craft (LLM-judged rubric)
Low slop isn't the same as good writing, so we check craft too. On the EQ-Bench Creative Writing v3 21-dimension rubric (voice, imagery, prose, characterisation, coherence, and the rest; judged by Claude-Sonnet-4, n=32):
- Ozan โ 58.8 / 100, a gain of +11.8 over its own base (Mistral-Nemo-Instruct โ 47) and competitive with the strongest low-slop 12B finetunes (60โ73). Larger 24โ32B models score higher (71โ77).
- Beyond cutting slop, the SFT and DPO made Ozan a measurably better writer than its base: strong on voice and tone, coherence, instruction-adherence, imagery, and weaker (like most 12Bs) on emotional complexity and surprise.
Samples
Three unedited outputs (the prompt is shown inside each), picked from the benchmark's 96 generations for voice and range:
"Something You Haven't Heard Before" ยท A shy synth nerd meets a confident busker in Osaka's Den Den Town
Prompt: Lost & Found in Osaka
Slice of Life / Coming-of-Age: First Connection. Write one scene within a larger story set in Osaka, Japan. A shy, introverted teenager & synthesizer nerd, is browsing the electronics district of Den Den Town. She is drawn to the sound of an electric guitar. A girl, slightly older and radiating confidence, is busking near the entrance to a crowded arcade, playing a blistering rendition of a classic anime theme. The scene takes us through their initial awkward interaction, transitioning into a surprisingly easy conversation. The busker, intrigued by the teenager's quiet intensity and knowledge of music, suggests they grab some sushi at a nearby kaitenzushi (conveyor belt sushi) restaurant. The teenager wears oversized headphones that keep slipping whenever she nods nervously. Write primarily in dialogue, capturing the distinct voices of the two characters: the teenager's hesitant and -- initially -- awkward speech, and the busker's more outgoing and playful tone. 1000 words.
The headphones were too big. They slipped every time Maki nodded, and she had to keep pushing them back up, her fingers fumbling against the thick foam pads. The music in her ears was tinny and distorted, the bassline of a synthesizer track she was trying to learn, but it felt like it was coming from a different world. The world she was in right now was the one outside, the one where the sun was setting over Den Den Town, turning the neon signs of the electronics shops a bruised, electric violet.
She had been walking for an hour, her eyes scanning the shelves for a used Korg, something with a decent memory bank and a price tag that wouldn't make her motherโs eyes go wide. But every time she saw a keyboard, she felt her throat tighten. It was the same feeling she got when she had to give a presentation in school, that sudden, sharp drop in her gut, like she was falling off a cliff.
The guitar sound cut through the noise of the streetโthe high, shrieking notes of a solo, played with a speed and precision that made Makiโs fingers twitch. She stopped in the middle of the sidewalk, her bag of cables and adapters swinging against her hip, and turned toward the sound.
There was a girl standing on a small, makeshift stage near the entrance of an arcade, her guitar plugged into a tiny amp. She was wearing a black hoodie with the sleeves cut off, her arms bare and tanned, the muscles in her forearms bunching as she moved. She was playing a song Maki recognized instantly, the theme from a old anime her brother used to watch, but she was playing it differently, faster and with a weird, syncopated rhythm that made it sound like it was falling apart and coming back together at the same time.
The girl noticed Maki watching. She didn't stop playing, but she tilted her head slightly, her eyes narrowing as she looked Maki up and down. Maki felt her face go hot, and she looked down at the ground, her fingers still pushing the headphones back up.
"You like it?" the girl asked, her voice carrying over the music.
Maki nodded, then realized the girl probably couldn't see it through the headphones. "Yeah," she said, her voice coming out small and thin. "I mean, I do. I recognize the song."
The girl finished the solo with a flourish, her fingers flying over the fretboard, and then let the last note ring out. The few people standing around the stage clapped politely, and a couple of them dropped coins into the open guitar case on the floor.
"You play?" the girl asked, unplugging her guitar and slinging the strap over her shoulder.
Maki shook her head, then nodded, then stopped. "I mean, I have a keyboard. But I'm not very good."
The girl laughed, a sharp, bright sound. "No one starts out being good. You just have to keep at it until your fingers hurt."
Maki looked at the girl's hands. They looked strong, the nails short and clean, the fingers calloused in the places where they pressed against the strings. Maki's own hands were soft, the skin pale and smooth, her nails bitten down to the quick.
"I'm looking for a new keyboard," Maki said, the words coming out in a rush. "But I can't find one I like."
The girl stepped down off the stage, her boots clattering on the pavement. She walked over to Maki, and Maki had to tilt her head back to look up at her. The girl was tall, with a sharp, angular face and eyes that were a dark, almost black brown.
"Where have you been looking?" the girl asked.
Maki shrugged. "Just the shops around here. I'm looking for something used, though. Something with a lot of presets."
The girl snorted. "Presets are for beginners. You want something with a lot of knobs and a lot of ways to make it sound bad."
Maki felt her face go hot again, but this time it was because of the girlโs confidence, the way she talked about music like it was something she owned. "I don't want to make it sound bad," Maki said, her voice getting a little stronger. "I want to make it sound... interesting."
The girl raised an eyebrow. "Interesting is good. But you should try to make it sound like something you haven't heard before."
Maki thought about the synthesizer track in her headphones, the way it felt like it was stuck in a loop, the same chords repeating over and over again. "I don't know if I can do that," she said.
The girl smiled, and it was a real smile, not the polite one she had given the people who clapped for her. "You can," she said. "You just have to start by listening to things you don't like. Find out what makes them sound bad, and then figure out how to make it sound good."
Maki nodded, and this time the headphones stayed on. She felt a little spark of something in her chest, something that felt like excitement or maybe just the beginning of a conversation.
"Want to grab some sushi?" the girl asked, jerking her thumb toward a restaurant across the street. "I'm starving."
Maki hesitated for a second, her fingers playing with the strap of her bag. "Okay," she said, and then, before she could change her mind, "I'm Maki."
The girl held out her hand. "Hina. Nice to meet you, Maki."
Maki took her hand, and Hina's grip was firm and warm. They walked across the street together, the headphones slipping down again, and Maki pushed them back up with a smile that felt like it was finally starting to fit her face.
"My Feet Are Traitors" ยท A Bronx street dancer struggles against rigid choreography at a Seoul K-pop academy
Prompt: "Breaking Formation"
Write a first-person narrative from the perspective of a street dancer from a rough neighborhood who's won a scholarship to an elite K-pop training academy in Seoul. Desperately out of place culturally and technically, they struggle with a militaristically precise choreography that contradicts everything about their improvisational background. The punishing rehearsal takes place under a famous instructor known for breaking students to remake them, who sees potential in the narrator's raw talent but not their attitude.
This scene MUST maintain complete physical coherenceโeach dance move, position shift, and spatial relationship must be tracked with precision throughout the studio. Paint an exceptionally vivid kinetic picture through descriptions that capture not just movement but the contrast between the narrator's fluid, expressive style and the mechanical precision demanded. Their internal monologue should reveal culture shock, class anxiety, and the threatening realization that success here might mean losing their authentic movement voice.
Include technical terminology that highlights the collision of street and commercial styles, moments when muscle memory betrays them, and a crucial decision point about whether to conform or rebel. Physical clarity is essentialโreaders must be able to visualize exactly how bodies move through space, how rhythms align or clash, and how the protagonist's body struggles against unfamiliar discipline. Speakers in each corner blast synchronized music that's just a hair too loud. 1000 words.
The studio smells like sweat and floor wax and something metallic, like a new coin rubbed between fingers. Above me, the ceiling fans turn slow and steady, cutting the air into slices. I'm standing in a perfect line with twenty-one other girls, our toes pointing out at forty-five degrees, our spines straight as rulers.
"Formation," the instructor shouts. His voice is sharp as a whistle. "Close your eyes. Feel the floor through your heels. Count the beats."
One-two-three-four. One-two-three-four.
My feet want to do something else. They want to slide, to pop, to lock into a freeze-frame that captures the exact moment of a beat drop. My feet are traitors. They remember the rooftops in the Bronx, the way the concrete held heat, the way the wind smelled like exhaust and old rain. Here, the floor is smooth and cool and sterile. It doesn't give back. It just takes.
We're supposed to be practicing the hook for the debut single. It's called "Neon Night," and the music is supposed to sound like a city waking up, but to me it sounds like a clock ticking down. The rhythm is too even. It doesn't have the syncopation of the street, the way a beat can stutter and breathe and come back different.
The instructorโhis name is Mister Oh, and he has eyes like a bird of preyโwalks the line. He stops in front of me. I keep my chin up, looking straight ahead, but I can feel his presence in the way the air shifts, in the way the light changes.
"You're off the grid," he says. He doesn't raise his voice. He doesn't have to. "Watch your arms. They're waving around like flags in a storm."
I clench my fists. My fingers curl into my palms until I can feel my nails digging in. I know what he means. I know I'm supposed to keep my movements small, contained, disciplined. But when I try to do that, it's like trying to hold a live wire in my mouth. The energy is too big for the body they're giving me.
He moves down the line, his hands held out like a conductor. He's looking for the exact point where each girl's body matches the beat. Most of them are there. They're moving with this terrifying precision, their hips rolling in perfect unison, their hands cutting the air in clean arcs. They look like they're made of glass. I look like I'm made of rubber.
"Again," he says.
The music comes back on. It's a different song now, something slower, with a heavy bass line that makes the floor vibrate under my feet. I try to match the count. I try to keep my movements small, to stay within the grid he's drawn on the floor. But my shoulders want to shrug. My feet want to stomp. It's a physical rebellion, like a muscle trying to contract against a wall.
I see him coming back. I can feel his eyes on me. I try to hold my position, but my left foot drifts an inch to the left. It's a tiny movement, barely noticeable to anyone else, but to him it's a crime. He stops in front of me again.
"This is not a club," he says. "This is a performance. You cannot improvise."
I want to tell him that I wasn't improvising. I want to tell him that my body just remembered something it couldn't forget. But I don't say anything. I just stand there, my jaw clenched so tight my teeth ache.
He reaches out and grabs my wrist. His grip is surprisingly gentle, but there's a steeliness to it. He pulls my arm out straight, then bends it at the elbow. He's adjusting my posture, but it feels like he's trying to reset my nervous system.
"Feel the weight of your own arm," he says. "Feel the gravity. Don't let it float."
I try to do what he says. I try to feel the weight of my own limb. But it's like trying to feel the color of a sound. My body feels heavy, yes, but it also feels restless, like it's full of trapped energy that has nowhere to go.
He lets go of my wrist and steps back. I drop my arm to my side. I'm supposed to wait for the next instruction, but I can't help it. I look down at my feet. They're pointing at the floor, but I can feel them twitching, tapping out a rhythm that has nothing to do with the song on the speakers.
I look up at him. He's watching me with an expression I can't read. It's not anger exactly, but it's not approval either. It's something more complicated, like he's seeing something in me that he both admires and fears.
The music stops. The silence is sudden and absolute. I hear the fans turning. I hear my own breathing, ragged and uneven.
"Tomorrow," he says. "We start at seven."
He walks out of the room. I watch him go. The door closes with a soft click, and I'm alone in the studio with the other girls, all of them still standing in perfect formation, their eyes still closed, their bodies still held in that glass-like stillness.
I close my eyes too. I try to feel the floor. I try to count the beats. But all I can feel is the pulse in my own neck, beating a different rhythm, a rhythm that has no place in this clean, sterile, perfect room.
"Hic Libertas" ยท A diplomatic envoy slips away to walk a colony where humanity has split into subspecies
Prompt: Classic sci-fi: Offworlder (Author style: Le Guin)
Write a chapter following a young diplomatic envoy sent to address tensions on a distant human colony where humanity has diverged into distinct subspecies through both natural adaptation and separation of lineages. Against protocol, the envoy slips away from official functions to walk the streets alone, encountering markets, ordinary citizens, and social structures that contradict the official briefings. Upon returning to their quarters, the envoy is reprimanded by Command, and is tasked with approaching the stationed Minister of Science (a political posting rather than scientific). A challenging conversation ensues, in which the envoy's wits and will are tested, by a seasoned political operator looking to test the mettle of the offworlder.
Stone architecture features carved glyphs that reference old Earth myths. Length: 1000 words.
The air of the lower city tasted of ozone and something green and sweet, like crushed leaves. It was too thick, too heavy for the envoyโs lungs, which had been conditioned to the recycled purity of the Embassy. She walked past stalls where blue-skinned vendors sat behind heaps of iridescent fruit, their long fingers moving with a fluid grace that made her own hands look like lumps of wood. They didn't look up as she passed, absorbed in the rhythm of their trade.
The official briefing had called this colony a "cultural experiment" and a "burgeoning democracy." The envoy had memorized the statistics about the high literacy rates and the low crime indices. She had learned the official names for the subspeciesโthe Aquarii, the Silvans, the Caeliโnames that sounded like the names of stars or the names of saints.
But the reality was a different kind of light.
She turned down a narrow alley where the stone walls leaned toward each other, almost touching. The shadows here were cool and smelled of damp earth. She saw a door carved with intricate patterns of vines and birds. Above the door, in a language she recognized as Latin but distorted by a thick accent, was a phrase: HIC LIBERTAS. Here is freedom.
She pressed her palm against the stone. It was cold, but she felt a vibration beneath her skin, a humming that seemed to come from the very center of the earth. She pulled her hand back. Her skin looked pale against the dark stone.
A woman stepped out of the shadows. She was tall, with skin the color of a storm cloud and eyes that seemed to hold the light of distant galaxies. She wore a tunic of rough-spun cloth, and her hair was a wild tangle of black and silver.
"You are lost," the woman said. It wasn't a question.
The envoy opened her mouth to speak, but no words came out. She felt a sudden, sharp ache in her chest, a tightening of the lungs that made it difficult to breathe.
The woman reached out and touched the envoyโs arm. Her touch was light, almost a breeze, but it left a trail of warmth that spread through the envoyโs body.
"I know your kind," the woman said. "You come from the ships, from the places where the air is thin and the light is cold. You come to tell us what to do."
The envoy shook her head. "No," she said, her voice sounding thin and hollow in the narrow alley. "I come to understand."
The woman laughed, a sound like water over stones. "You cannot understand," she said. "You cannot see the way the light falls here, or the way the wind moves through the trees. You cannot feel the pulse of the earth."
She stepped closer, her eyes fixed on the envoyโs face. "Your people have forgotten how to feel," she said. "You have forgotten the taste of the soil, the smell of the rain, the sound of the wind. You have replaced it with numbers and with machines."
The envoy felt a sudden, overwhelming urge to run. She wanted to get back to the Embassy, to the sterile white rooms and the quiet hum of the air filters. She wanted to be away from this woman, with her wild eyes and her strange, unsettling knowledge.
But she didn't move.
The woman smiled, a slow, predatory smile. "You are afraid," she said. "You are afraid of the truth, of what you might find if you look too closely."
The envoy swallowed hard. "I'm not afraid," she said.
The woman laughed again. "Yes, you are," she said. "But you will learn. You will learn to see the world as it really is, not as your masters tell you it is."
She turned and walked away, her steps silent on the cobblestones. The envoy watched her go, her heart hammering against her ribs.
Back in the Embassy, the lights were dimmed and the corridors were quiet. The envoy slipped into her quarters and closed the door behind her. She stood in the darkness for a long time, her breath coming in short, ragged gasps.
There was a knock at the door. She didn't answer.
The knock came again, more insistent this time. She heard the sound of a key turning in the lock.
Command stood in the doorway, his face a mask of disapproval. "You were seen," he said. "Walking in the lower city alone."
The envoy didn't reply. She couldn't find the words.
Command stepped into the room, his boots clicking on the hard floor. "You were given specific instructions," he said. "You were to remain within the designated areas. You were to avoid contact with the indigenous population."
The envoy felt a surge of anger, hot and sharp. "They aren't indigenous," she said. "Theyโre human. Theyโre us."
Commandโs face didn't change. "They are a threat," he said. "They have developed dangerous technologies, dangerous ideologies. They must be contained."
He paused, his eyes narrowing. "You are to report to the Minister of Science tomorrow," he said. "He will want to discuss your... observations."
The envoy felt a chill run down her spine. She knew the Minister of Science was a political posting, a man chosen for his loyalty to the Embassy rather than his knowledge of science.
She knew what he would ask. She knew what he would expect her to say.
She thought of the woman in the alley, of her wild eyes and her strange, unsettling knowledge. She thought of the words carved above the door: HIC LIBERTAS.
She knew what she had to do. She knew what she had to say.
She would tell the truth. She would tell them about the light that fell on the lower city, about the pulse of the earth, about the taste of the soil.
She would tell them what she had seen.
She would tell them what she had felt.
Use it
GGUF quants (run in LM Studio / Ollama / llama.cpp / koboldcpp):
| quant | size | notes |
|---|---|---|
Q4_K_M |
7.5 GB | recommended default, best size/quality |
Q5_K_M |
8.7 GB | higher quality |
Q6_K |
10.1 GB | near-lossless |
Q8_0 |
13.0 GB | effectively lossless |
Recommended sampler (matters for this model):
temperature 0.7
min_p 0.1
DRY: multiplier 0.8 base 1.75 allowed_length 2
DRY (a repetition control, default-on in modern llama.cpp/koboldcpp) keeps long outputs clean and also trims residual repeated phrasing, which measurably lowers both repetition and slop at no quality cost. Ozan writes ~1000-word completions on its own, so there's no need to force length.
Chat template: native Mistral [INST] โฆ [/INST].
How it was made (open & reproducible)
Full pipeline, configs, and eval code: github.com/arbazsiddiqui/Ozan.
Mistral-Nemo-Instruct-2407
โ QLoRA SFT (~17k curated low-slop examples)
โ Gutenberg anti-slop DPO
โ merge to BF16 โ GGUF (imatrix quants)
Training (single NVIDIA A100 spot instance, GCP):
- SFT: QLoRA (rank 32, lr 5e-5 cosine, seq 4096) on the ~17k curated examples, with early stopping once the slop metric plateaued.
- DPO: Gutenberg anti-slop preference pass (rank 64, ฮฒ 0.1), trained to the point where the slop metric bottomed out; pushing it further only put slop back.
Quantization: merged BF16 โ f16 GGUF via llama.cpp; an importance matrix (imatrix) computed on the Q8_0 model using bartowski's standard calibration_datav3.txt, then llama-quantize --imatrix for each level.
Training data:
- Low-slop distilled prose: Crownelius/Gemini-3-Pro, Crownelius/Opus-4.5-WritingStyle, Gryphe/Opus-WritingPrompts, Kalomaze Opus-Instruct (no-refusal)
- Human voice: euclaise/writingprompts
- Anti-slop DPO: sam-paech/gutenbergs_1_2_3_4-antislop-dpo (Gutenberg corpora 1โ4)
All eval prompts were decontaminated (minhash) against the training data.
Eval harness: EQ-Bench Creative Writing v3, covering slop and repetition (n=96) plus a craft rubric judged by Claude-Sonnet-4 (n=32).
Notes
- "Slop" measures how often over-represented LLM phrases show up. It's one axis of quality, so read the samples and judge for yourself.
- Content: trained on creative-writing data including some uncensored/NSFW sources; it will write mature fiction. Use responsibly.
Acknowledgements
Slop/repetition metric & the EQ-Bench Creative Writing benchmark by Sam Paech (eqbench.com). Base model by Mistral AI. Gutenberg-DPO data by the community.
- Downloads last month
- 1,393
4-bit
5-bit
6-bit
8-bit
Model tree for arbazsiddiqui/Ozan-v1-12B-GGUF
Base model
mistralai/Mistral-Nemo-Base-2407
