Bazza82/barry-li-style-deai-editor-qwen3-8b-lora

Barry Li Style De-AI Editor Qwen3-8B LoRA

This repository contains a LoRA adapter for unsloth/Qwen3-8B-bnb-4bit.

The adapter is designed to edit draft prose into a clearer Barry Li style: direct, reflective, practical, and grounded in human judgement. It is intended for users who want to reduce generic machine-like writing and make prose read with a more natural authorial voice.

It is not a full model. Users need the base model plus this adapter.

About Barry Li

Barry Li is an AI and technology transformation leader based in Sydney, Australia. His work sits across agentic AI, enterprise data platforms, digital transformation, technology governance, audit and assurance, and organisational change.

Barry began his career in audit and assurance before moving into technology-enabled transformation, data capability uplift, AI-enabled workflows, secure data use, and scalable approaches to complex knowledge work. He is a CPA, a PhD candidate in accounting and finance, a published author, and a multilingual professional working across English, Mandarin Chinese and Japanese-speaking contexts.

Barry is the author of The New Chinese: How They Are Shaping Australia, published by Wiley in 2017. This adapter was trained to reflect aspects of Barry Li's published writing style, including material from The New Chinese, together with curated writing-style edit examples.

Intended Use

Use this adapter as a writing-style editor. It can help revise draft academic, professional, or explanatory prose so that it becomes clearer, less generic, and closer to Barry Li's reflective nonfiction voice.

The adapter is most useful when the source text is already broadly meaningful but reads as overly mechanical, inflated, impersonal, or over-compressed. It aims to improve voice, clarity, sentence movement, and readability rather than to verify facts or rewrite content from first principles.

Not Intended For

This adapter is not designed for:

• factual verification
• preserving citations perfectly
• legal, medical, financial, or safety-critical advice
• producing final academic text without human review

The model may alter citations, terminology, dates, names, or technical details. Human review is required.

Recommended Prompt

/no_think. Rewrite the source text in Barry Li's writing style.
Make the prose clearer, more direct, and more human.
Prefer plain verbs, clear judgement, and reflective nonfiction rhythm.
Keep the broad meaning, but prioritise voice and readability over strict academic formality.
Do not explain the edit. Do not list changes. Write only the revised passage.

Source text:
{your text here}

Example

Source

This paragraph explains the methodological value of using computational text analysis in a literature review, but the original draft is overly dense and impersonal.

Output

The value of computational text analysis is not that it replaces the researcher. Its value is more practical. It gives the review a way to work through a large and changing literature while keeping the judgement, interpretation and responsibility with the researcher.

Loading With Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "unsloth/Qwen3-8B-bnb-4bit"
adapter = "Bazza82/barry-li-style-deai-editor-qwen3-8b-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)

prompt = """Rewrite the source text in Barry Li's writing style.
Write only the revised passage.

Source text:
This paragraph explains the methodological value of using computational text analysis in a literature review, but the original draft is overly dense and impersonal."""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=250, temperature=0.6, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Summary

• Base model: unsloth/Qwen3-8B-bnb-4bit
• Method: QLoRA / LoRA adapter
• Train records: 202
• Eval records: 32
• Epochs: 2
• Final train loss: about 0.2874
• Eval loss: about 0.5473

The training data included curated style-edit examples and Barry Li style anchors, including material derived from Barry Li's published book The New Chinese: How They Are Shaping Australia (Wiley, 2017). The raw manuscript and training data are not included in this repository.

Known Issues

• May introduce typos.
• May alter citations or technical terms.
• May over-personalise academic text.
• Requires human review before publication.

Suggested Citation

Barry Li Style De-AI Editor Qwen3-8B LoRA, 2026.
LoRA adapter for Qwen3-8B designed to edit prose into Barry Li's writing style.