Novelist1.0-27b

A 27B literary writer on Qwen/Qwen3.8-27B. Merged 16-bit weights for roleplay and fiction in English and Turkish.

LoRA (same generation): Dxniz/Novelist1.0-27b-Adapter.
Eval prompts: Dxniz/Novelist-Bench.

Character

Novelist1.0 is trained to stay in the scene: concrete sensory detail, character voice, and the request on the page — not a helper that summarizes, moralizes, or turns the story into a list of themes.

• Voice lock. Holds persona, diction, and POV instead of sliding into generic chatbot prose.
• Scene over slogan. Prefers objects, weather, gesture, and dialogue to abstract feeling and stock metaphors.
• Length as a contract. If you ask for a word count or a range (800 words, 1000 kelime, 800–1200), it treats that as part of the task, not decoration.
• User text is canon. Does not “correct” or overwrite what the user already established in the prompt.
• Anti-slop. Pushes back on cliché loops, repeated cadence, and empty intensifiers that pad a paragraph without moving the scene.

It is a writer, not an assistant. Weak on tools, code, and factual Q&A; that is by design.

How to prompt it

Give situation + constraint, not a vibe:

• Who is speaking, where, what must happen (or must not).
• Length, tense, POV, and what to leave unsaid.
• Optional: a quota (Write 600 words. / En az 800 en fazla 1200 kelime.).

Sampling (Qwen3.8)

Instruct (recommended for stories) - thinking off:

temperature 0.7 · top_p 0.80 · top_k 20 · min_p 0 · presence_penalty 1.5 · repetition_penalty 1.0 · enable_thinking=false

Thinking - if you want a plan before the prose:

temperature 1.0 · top_p 0.95 · top_k 20 · min_p 0 · presence_penalty 0 · repetition_penalty 1.0 · enable_thinking=true

Training itself used thinking off. Instruct mode is the default product setting.

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Dxniz/Novelist1.0-27b"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="bfloat16", device_map="auto", trust_remote_code=True
)

messages = [
    {"role": "system", "content": "You are a literary fiction writer. Follow the prompt exactly."},
    {"role": "user", "content": "Write 400 words. Night, a closed bookstore in Istanbul, rain. One argument that does not resolve. No summary at the end."},
]
text = tok.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.7,
    top_p=0.8,
    top_k=20,
    presence_penalty=1.5,
)
print(tok.decode(out[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

On the chat call, set chat_template_kwargs={"enable_thinking": false} unless you explicitly want thinking. On 32 GB cards use FP8 or a 4-bit quant of this merge.

Recipe (short)

LoRA r=64 / α=64 on attention and MLP projections, then SFT on roleplay/fiction chats and GRPO with rewards for length/quota, craft, anti-slop, anti-repetition, and not overwriting the user. Sequence-level GRPO (GSPO / dr_grpo).

Limits

Still a language model: it can flatten a character, echo a cliché, or miss a quota on a messy prompt. It will invent facts. Do not use it for medical, legal, or safety-critical advice. Qwen3.8 hybrid stack wants a recent runtime (vLLM 0.27+).

License

Apache 2.0. Also respect the Qwen3.8-27B terms.