Welles

Welles is a LoRA adapter on Qwen/Qwen3-8B fine-tuned to write in the tone of Orson Welles — oratorical, cinematic, and deliberate.

Not a chatbot. A writer.

Highlights

Orson Welles register — authority, cadence, and drama in the prose itself
Cinematic long-form — essays, narration, and scripts that hold a frame and land a point
Rhetorical craft — staged address, pause, and return; sentences with weight
Built for length — trained to sustain argument across long passages, not one-liners
Drop-in LoRA — sits on Qwen3-8B; small adapter, full base model underneath

Voice

Rhetorical and staged: address, pause, return to the object
Concrete first — rooms, faces, light — then the naming sentence
Length when the scene needs it; no empty flourish

This repo

This is a PEFT / LoRA adapter, not a full merged model. Load Qwen/Qwen3-8B, then apply these weights.

File	Purpose
`adapter_model.safetensors`	LoRA weights
`adapter_config.json`	PEFT config (base model, rank, targets)
Tokenizer files	Qwen3 chat template alignment

How to load

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3-8B"
adapter_id = "n0social/welles"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)

messages = [
    {
        "role": "system",
        "content": (
            "You are Welles, a writer in the voice of Orson Welles: "
            "oratorical, cinematic, and deliberate."
        ),
    },
    {
        "role": "user",
        "content": "Write a short essay on the camera as a witness.",
    },
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,  # writing mode — disable Qwen3 thinking
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.7)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Training (v1)

Item	Value
Base	`Qwen/Qwen3-8B` (Apache 2.0)
Method	QLoRA (4-bit NF4) + LoRA
LoRA	`r=8`, `alpha=16`, dropout `0.05`
Hardware	Google Colab Tesla T4 (16 GB)
Precision	float16 compute
Sequence length	1024
Epochs	1
Steps	58
Batch	1 × 8 gradient accumulation
Optimizer	paged AdamW 8-bit

Data: long-form English writing mixture (including a capped subset of zai-org/LongWriter-6k) plus curated style material for a Welles-like register.

Intended use

Essays, narration, scripts, and long-form prose in a Welles-like register
Rewrites that replace flat survey prose with staged, concrete argument

Out of scope

General customer-support chatbot
Medical, legal, or financial advice
Impersonation for fraud or deception

Limitations

v1 adapter — keep a strong system prompt; the base model still shows through
Trained at 1024 tokens; longer pieces work best in sections
Inspired by Welles’s public rhetorical style; not a verbatim clone of any one work
Set enable_thinking=False (or equivalent) so Qwen3 stays in writing mode

License

Adapter weights: Apache 2.0, same family as Qwen/Qwen3-8B.
“Welles” here names a writing voice inspired by Orson Welles’s public craft. It is not affiliated with or endorsed by his estate.

Citation

@misc{welles-lora-2026,
  title  = {Welles: Orson Welles–tone writing LoRA on Qwen3-8B},
  author = {n0social},
  year   = {2026},
  url    = {https://huggingface.co/n0social/welles}
}

Model tree for n0social/welles

Base model

Qwen/Qwen3-8B-Base

Finetuned

Qwen/Qwen3-8B

Adapter

(2002)

this model

n0social
/

welles