← All cases Learning project

Fine-tuning a local LLM to write in a brand's voice

I fine-tuned the open Vikhr-Llama model with LoRA so it writes in a given brand's tone of voice. It runs locally on ordinary hardware, and the data never goes to the cloud.

Task

A brand with its own voice needs a steady flow of texts that sound like the brand. Cloud models can help, but then every draft and every piece of data goes to someone else's servers. I built this as a learning project to try the other route: fine-tune an open language model on a brand's style and run it locally, with no cloud dependencies.

Solution

I took Vikhr-Llama, an open language model, and fine-tuned it with LoRA. In short, LoRA trains a small add-on to the model instead of retraining all of it. A brand's tone of voice ends up in the model itself, so you don't have to spell it out in every prompt. The model runs locally on ordinary hardware, and the data doesn't leave your machine.

Results

  • ✓ Content comes out 3–5x faster
  • ✓ A brand's tone of voice is built into the model itself
  • ✓ Full privacy: the model runs locally, data stays off the cloud