LLM fine-tuning on your examples, terminology and tone of voice
We train an open-weight model on your conversations, documents and approved answers. It starts using your terms and formats, and we measure quality on a test set before and after.
What hurts
The model gives generic answers and does not know your terminology. The prompt is already several pages long, and it still mixes up products and uses the wrong tone.
You pay for a large model where the task is narrow and repetitive. Call review or ticket classification could run on a smaller model if it were trained for the job.
Someone pitches fine-tuning as a magic button. But nobody can tell you how to check that the model got better after training rather than worse.
What we build
- Task analysis: whether fine-tuning is needed or RAG and prompting will do
- Collecting and cleaning training examples from your data
- Removing personal data before training, in line with GDPR
- A test set and quality metrics agreed before training starts
- LoRA fine-tuning on one or two base models
- Comparison of the tuned model with the base model and a cloud API
- Deployment of the tuned model on vLLM
What you get
- Weights of the fine-tuned model, owned by you
- The cleaned training set and the scripts that prepare it
- A test set with before and after results
- A deployed model with an API, or a deployment guide
- A retraining procedure for when new data arrives
How we measure the result
- Test set score against the base model
- Share of answers in the correct format without manual edits
- Cost per request against a large cloud model
- Number of errors in terms and product names
Systems and stack
Industries
Questions
How much does LLM fine-tuning cost?
From $10,000. The price depends on the volume and cleanliness of data, the number of tasks and the model size. The most expensive part is usually not training itself but preparing examples and evals. GPU time for training is included in the estimate.
How much data do we need?
A narrow task often works with a few hundred good examples. For style and terminology a few thousand is better. Correctness matters more than volume. During discovery we look at your data and tell you honestly whether it is enough.
Is fine-tuning always better than RAG?
No. If the model must know fresh facts, prices or stock, RAG is better. Fine-tuning gives format, style and behaviour. A combination often works best. We recommend fine-tuning only when the evals show a gain.
Can our data leak through the model?
We remove personal data before training and train inside a closed environment. A model can memorise fragments of its training data, so sensitive fields never enter the set. We sign an NDA and DPA before access.
What happens after launch?
The model ages along with your business. We set up a retraining procedure and can run it under an SLA. Every new version passes the same test set before it replaces the old one.
Which base model do you start from?
Open-weight models whose licence allows commercial use and fine-tuning. We pick two candidates by size, language and speed, tune both on part of the data and keep the one that scores better. We check Polish, German or other EU languages separately, because models handle them unevenly.
Related services
More in this section
Describe the task. We reply within one business day, the first call is with the founder.
Describe the task. The first call is with the founder.
We reply within one business day. 30 minutes to tell you whether it is worth doing and what it costs.