Skip to content
Part of Agent & LLM Consulting →

LLM fine-tuning, when your evaluation shows it pays.

We tune open-weight models on your examples with the method your data supports, and score the result against the model it started from.

Tuning changes behaviour. Retrieval supplies facts.

Fine-tuning teaches a format, a tone or a decision pattern. Facts that change belong in retrieval, where they can be updated the same day.

What does not change

Data first

The examples are checked for quality, coverage and overlap with the test set before any training starts.

Measured against the base model

Every tuned model is scored on the same held-out set as the model it started from.

Yours to keep

The adapter or the tuned weights, the training data and the scripts, within the terms of the base model's licence.

The methods, and what each needs

Supervised fine-tuning

Pairs of an input and the answer you want. The usual first step, and the one most projects need.

LoRA and QLoRA

Small adapter weights trained on top of a frozen model. Cheaper to train and to store; QLoRA fits larger models on smaller GPUs.

Preference tuning

Pairs of a better and a worse answer, as in DPO. It teaches which of two acceptable answers you prefer.

Reinforcement-style tuning

Rewards from a checker, such as a test suite or a rule, as in GRPO. It suits tasks with a right answer a program can verify.

Distillation

A large model's answers train a smaller one, which then runs faster and cheaper on the narrow task.

Continued pre-training

A large body of unlabelled text in a language or field the base model knows poorly.

Questions buyers ask

How much data do we need?

That depends on the method and the task. We look at what you have and tell you what it supports before any training starts.

Can you fine-tune Claude or GPT for us?

Not their current models: neither vendor offers it today. We tune an open-weight model, such as gpt-oss, Qwen or Llama, and you keep the result.

Where does training run?

On your GPUs, on NVIDIA GPUs rented from RunPod or Lambda, or with NVIDIA NeMo, inside the boundary your data policy sets.

Last reviewed:

Bring examples of the work done right.

We check what they support and whether tuning beats the base model on your task.

Talk to us