yuxinlu1/qwen3-6-27b-chinese-crime-fiction-lora-v2

📚 Qwen3.6-27B Chinese Crime Fiction LoRA — v2

A LoRA adapter that guides Qwen3.6-27B toward Chinese suspense and crime-fiction prose with stronger control over narrative behavior, POV discipline, and scene-based drafting.

This adapter is designed for a specific writing problem:

Given a short instruction or an exposition-heavy setup paragraph, the model should directly produce usable Chinese fiction prose rather than analysis, outline, summary, or revision commentary.

The core direction is intentionally narrow:

• third-person limited narration
• stronger POV discipline and reduced perspective drift
• less exposition-heavy character introduction
• fewer omniscient reveals and narrator shortcuts
• direct fiction prose instead of analysis, outlines, or revision advice
• scene progression through action, sensory detail, objects, clues, and environment
• crime, investigation, forensic, suspect, witness, interview, and urban suspense scenarios

The goal is for outputs to behave more like usable Chinese suspense-crime scene drafts, and less like a generic writing assistant explaining how a scene should be written.

Typical use cases:

• drafting suspense or crime-fiction scenes inside a longer novel project
• rewriting exposition-heavy paragraphs into scene-based prose
• testing third-person-limited POV behavior
• local and private creative-writing workflows
• integration with a long-form novel pipeline that manages outline, memory, timeline, and continuity

This is an adapter only. It does not include base model weights, training data, or copyrighted source material.

---

🧭 About the v1 / v2 Series

This repository is part of a small series of Chinese fiction LoRAs.

v1 — Style Retraining

The v1 series focuses mainly on prose style.

It uses SFT to move the base model away from generic AI prose and toward specific Chinese literary style directions. Different v1 adapters may target different prose voices, such as realistic fiction, literary prose, or other narrative styles.

v2 — Style + Behavioral Fine-Tuning

• Adds DPO training, 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.
• Available formats: HF PEFT safetensors and GGUF LoRA.
• MLX users may also be able to use the PEFT safetensors through mlx-lm depending on their local setup.

v2 models are documented in their own repositories.

---

🌱 Status

Field	Value
Version	v2
Focus	Chinese suspense / crime-fiction behavior
Format	HF PEFT safetensors + GGUF LoRA
Base model	Qwen3.6-27B
Language	Chinese
Use case	fiction drafting, scene rewriting, POV-controlled suspense prose
Training style	SFT + manual DPO refinement
Recommended workflow	local long-form writing pipeline

---

🔗 Companion Novel Pipeline

This LoRA is designed to work together with a local-first, long-form Chinese novel writing pipeline:

github.com/DuckTraDo/Novel

The pipeline handles the structural side of long-form fiction:

• outline and chapter planning
• scene-level context assembly
• story memory and character tracking
• timeline and continuity checks

This LoRA handles the prose-behavior side:

• POV discipline (third-person limited)
• anti-exposition and anti-meta-writing
• scene-based suspense / crime-fiction drafting

They are intentionally split: the pipeline owns what happens, the LoRA owns how it reads on the page. You can use either independently, but they are designed as a pair.

---

📦 Files

This repository provides the LoRA adapter in both HF PEFT and GGUF LoRA formats:

adapter_config.json
adapter_model.safetensors
qwen3-6-27b-chinese-crime-fiction-lora-v2-f16.gguf
tokenizer.json
tokenizer_config.json
chat_template.jinja
README.md

This is not a full model.

The GGUF file is a LoRA adapter for llama.cpp-style inference. It is not a merged full model.

To use it, load the Qwen3.6-27B base model and then apply this adapter.

---

🚀 Example PEFT / Transformers Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "Qwen/Qwen3.6-27B"
adapter = "yuxinlu1/qwen3-6-27b-chinese-crime-fiction-lora-v2"

tokenizer = AutoTokenizer.from_pretrained(
    base_model,
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)

model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [
    {
        "role": "user",
        "content": "写一段刑警深夜重返案发现场的悬疑犯罪小说。"
    }
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=800,
    temperature=0.7,
    top_p=0.85,
    do_sample=True,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

🚀 Example llama.cpp Usage

Use a compatible Qwen3.6-27B GGUF base model, then load this GGUF LoRA adapter with --lora:

llama-server \
  -m path/to/qwen3.6-27b-base.gguf \
  --lora qwen3-6-27b-chinese-crime-fiction-lora-v2-f16.gguf \
  --ctx-size 16384 \
  --n-gpu-layers 99 \
  --port 18084

For Windows PowerShell:

C:\llama.cpp\llama-server.exe `
  -m "C:\path\to\qwen3.6-27b-base.gguf" `
  --lora "C:\path\to\qwen3-6-27b-chinese-crime-fiction-lora-v2-f16.gguf" `
  --ctx-size 16384 `
  --n-gpu-layers 99 `
  --port 18084

---

🧪 Example Prompts

This adapter was trained on a mixture of detailed prompts, short prompts, and minimal one-line instructions. It does not require a long system prompt to start writing fiction prose.

Example prompts:

写一段刑警夜里重返案发现场的悬疑犯罪小说。

小说：旧书店地下室，凌晨。

写一段关于纪实作家查案的小说。

把下面这段改成悬疑犯罪小说正文：

公益律师朱向北在市医院急诊走廊发现了异常：一只没有落灰的相框，一段被剪掉的录音，以及一个始终避开监控的嫌疑人。

The adapter is designed to respond to ordinary fiction-writing instructions and produce Chinese prose directly.

Recommended decoding range:

temperature: 0.6–0.8
top_p:       0.8–0.9
max_new_tokens: 600–1200

---

🧪 Internal Evaluation Snapshot

A small 30-prompt internal evaluation was used to track behavioral progress during training.

Behavioral progression (SFT → final v2):

Metric	SFT baseline	Final v2
Exposition-pattern hits	6	3 (−50%)
Clean prose outputs	28 / 30	28 / 30
Meta / explanation leakage	2 / 30	2 / 30
Dialogue-required prompts	3 / 3	3 / 3
Average output length	~401	~436

The behavioral fine-tuning stage cut exposition-style writing (character-label dumping, omniscient narrator shortcuts) roughly in half while keeping clean-output rate stable.

These numbers come from a small local development eval and should be treated as an internal signal, not a public benchmark.

---

🧪 Intended Use

This LoRA is intended for:

• local Chinese suspense and crime-fiction drafting
• rewriting exposition-heavy paragraphs into scene-based prose
• testing third-person-limited narration
• creative writing assistance in a human-in-the-loop workflow
• offline and privacy-respecting novel drafting
• integration with a long-form novel pipeline

It is not intended for:

• impersonating any specific living author
• generating defamatory or harmful content about real people
• replacing human authorship, editing, or legal review
• mass-producing low-quality spam or content-farm material
• any use that violates the base model license or local laws

---

⚠️ Limitations

This adapter focuses on scene-level prose behavior. Long-form story structure still works best when guided by an outline, memory system, or human author.

Known limitations:

• It does not guarantee full-novel plot coherence by itself.
• Character continuity, foreshadowing, and timeline logic should be handled by an external writing pipeline or by the author.
• It may still occasionally produce revision-style commentary under some sampling settings.
• It may produce shorter-than-expected outputs if the prompt is very minimal or decoding settings are conservative.
• It is optimized for suspense and crime-fiction scenes rather than romance, fantasy, comedy, or web-novel action pacing.
• Output quality depends on the base model, quantization, sampling settings, prompt design, and context quality.

---

🛡️ 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 fictional. Any resemblance to real people or events is coincidental.
• Users are responsible for following applicable laws, platform policies, and the base model license when generating, storing, or publishing outputs.

---

📜 License

• LoRA adapter: MIT
• Base model: governed by its own license. Please check the upstream Qwen3.6-27B license terms before combining this adapter with the base weights.

You are responsible for complying with the base model's license when using this adapter.