Looking for a local LLM workflow that can learn my writing style from my published papers

Posted by Interesting_Pea_4404@reddit | LocalLLaMA | View on Reddit | 5 comments

Hi everyone, I am looking for a local LLM setup that can adapt to my writing style using my published papers as reference material.

I want to run it privately on my own machine (GeForce RTX 3080 ) to help with revising manuscripts and emails in an academic style. My papers are in a technical academic writing style, so I care most about preserving tone, structure, clarity, and terminology rather than general creativity.

I would appreciate recommendations for local models, fine-tuning or LoRA tools, and whether RAG might be better than fine-tuning for this use case.

Thanks.

[-]

loniks@reddit

For style — LoRA fine-tune beats RAG. RAG retrieves chunks but won't internalize your tone.

Try Mistral 7B or Qwen2.5 7B with unsloth (QLoRA 4-bit) — fits on a 3080, fine-tunes in under an hour. Convert papers to text with Marker, structure as instruction pairs.

Where RAG helps: terminology consistency. Keep a small index so the model uses your exact terms. But style should come from weights, not retrieval.

What field are you in?

Interesting_Pea_4404@reddit (OP)

Thank you for your detailed response. Seems interesting, I am in the civil engineering field.
How many papers do I need to fine-tune it?