📚 Qwen3.6-27B Chinese Realistic Fiction LoRA — v1
A LoRA adapter that pushes Qwen3.6-27B toward a Chinese realistic fiction prose voice for local long-form novel drafting.
The style direction is intentionally narrow:
- restrained, cold, plain narration
- dark humor
- working-class / everyday-life subject matter
- emotional understatement, silence, and negative space
The goal is for outputs to read more like a quiet realist short story in Chinese, and less like a web novel or a "polished AI essay."
Typical use cases:
- drafting scenes inside a longer novel project
- experimenting with literary style on a local model
- offline / privacy-respecting creative writing
This is an adapter only. It does not include base model weights, training data, or any copyrighted source material.
🧭 About the v1 / v2 Series
This repo is part of a small series of Chinese fiction LoRAs. The series is split into two generations.
v1 — Style retraining
- Trained with SFT to move the base model away from generic AI prose and toward a specific Chinese literary style.
- Within the v1 series there will be multiple LoRAs in different style directions. This repo is the realistic fiction direction.
- Format: GGUF LoRA (for
llama.cpp/ local inference).
v2 — Style + behavioral fine-tuning
- Adds DPO, larger amounts of synthetic data, and on-policy sampling on top of the v1 recipe.
- Output is more precise, more stable, and shows fewer habitual "AI-shaped" patterns.
- Released in multiple formats: HF PEFT safetensors / GGUF LoRA / MLX LoRA.
This repo is a v1 (style) adapter. v2 models are documented in their own model cards.
🌱 Status
| Field | Value |
|---|---|
| Version | v1 (style) |
| Style | Chinese realistic fiction |
| Format | GGUF LoRA |
| Base model | Qwen3.6-27B |
| Language | Chinese |
| Use case | fiction drafting |
🚀 Example llama.cpp Usage
llama-server.exe ^
-m path\to\your\base-model.gguf ^
--lora path\to\qwen3_6_27b_novel_lora.gguf ^
--port 18084 ^
--ctx-size 16384 ^
--n-gpu-layers 99 ^
--cache-type-k q4_0 ^
--cache-type-v q4_0 ^
--flash-attn
Then point an OpenAI-compatible client (or the Novel pipeline) at the local server:
$env:LLM_BASE_URL="http://localhost:18084/v1"
$env:LLM_API_KEY="local"
$env:LLM_MODEL="your-model-name.gguf"
🧪 Intended Use
This LoRA is intended for:
- local long-form Chinese fiction drafting
- experimenting with restrained, literary prose styles
- creative writing assistance in a human-in-the-loop pipeline
- offline / privacy-respecting writing workflows
It is not intended for:
- impersonating any specific living author
- generating defamatory or harmful content about real people
- mass-producing low-quality content for spam or content farms
- any use that violates the base model's license or local laws
⚠️ Limitations
- This is a v1 style adapter. It changes prose voice; it does not guarantee plot coherence or long-context consistency.
- Long-form continuity (characters, foreshadowing, timeline) needs to come from a writing pipeline or from the author, not from the LoRA.
- Output quality depends heavily on the base model, quantization, prompt, and outline quality.
- Style strength varies across sampling settings and hardware.
- The LoRA is narrow on purpose: if you ask it to write romance, fantasy, or web-novel-style action, results will be off-style.
🛡️ Safety and Legal Notes
- No copyrighted novels, private manuscripts, or proprietary datasets are distributed in this repository.
- This LoRA is not designed to imitate any specific living author.
- Generated text is fiction. Any resemblance to real people or events is coincidental.
- Users are responsible for following applicable laws and the base model's license when generating, storing, or publishing output.
📜 License
- LoRA adapter: MIT
- Base model: governed by its own license (see Qwen3.6-27B upstream terms).
You are responsible for complying with the base model's license when combining this LoRA with the base weights.
🔗 Related Project
This LoRA was built with a local-first novel writing pipeline in mind:
- Novel pipeline: https://github.com/DuckTraDo/Novel
The pipeline handles outlines, scene-level context, story memory, and continuity checks; this LoRA handles prose style. They are designed to be used together.