sus-tech

LLM Fine-Tuning with LoRA

Some tasks need a model that knows your vocabulary, your document formats or your way of answering. We fine-tune open language models with LoRA and QLoRA, which train small adapter layers instead of the whole model. That needs far less GPU memory and time. The result can run on your own hardware.

LLM Fine-Tuning with LoRA

When fine-tuning is the right tool

Fine-tuning changes how a model writes and decides. It fits when the model should follow a fixed format, use your terminology or handle a narrow task very reliably, such as classifying tickets or extracting fields from documents. When the model mainly needs current facts from your documents, a RAG knowledge base is usually the better choice.

How a project runs

  1. We define the task and how success is measured.
  2. We prepare training data from your examples and check it for quality and personal data.
  3. We train a LoRA or QLoRA adapter on an open model such as Llama, Mistral or Qwen, with open tools like Hugging Face PEFT or Unsloth.
  4. We compare the result with the base model on your own test cases.
  5. The model goes into your workflows, on your own hardware or in an EU data centre. vLLM serves the adapter together with its base model; several adapters can share one base model.

Why LoRA and QLoRA

LoRA trains a few million parameters instead of billions. QLoRA also loads the base model in 4-bit precision, so larger models fit on a single GPU. Our post LLM fine-tuning with LoRA and QLoRA explains both methods.

Ready to get started?

Tell us which process you want to improve. We look at it with you and suggest how to start.

Request consultingMore solutions