Capable, but it does not know your terms, formats or tone.
We collect good past work: reports, replies, records.
The model learns from your examples and follows them closely.
We compare it with the general model on real cases before use.
LLM Fine-Tuning
Language models trained on your examples.
Fine-tuning means further training an existing language model on your own examples, so it follows your formats, terms and style more closely. We advise when it helps and build it properly when it does.
What you get
01Honest advice firstA clear view on whether fine-tuning, better prompts or document search suits you.
02A model that fitsOutput that follows your formats, terminology and tone more consistently.
03Measured resultsSide-by-side tests showing how the tuned model compares with the original.
How we deliver
Four stages, each ending with something you can see and review before the next one starts.
01
Baseline
We measure how well existing models do on your task, so gains can be proven.
02
Curate
We build a clean, balanced training set from your examples, with your experts reviewing it.
03
Train
We fine-tune and compare several versions against the baseline on held-back tests.
04
Release
We deploy the best version, watch its output and retrain when your needs shift.
What we buildOur LLM fine-tuning capabilities
Fine-tuning assessment
We test simpler options first and recommend fine-tuning only when it clearly helps.
Discuss thisTraining data preparation
Collecting, cleaning and formatting your examples into a reliable training set.
Discuss thisSupervised fine-tuning
Training a model on example questions and ideal answers from your own work.
Discuss thisEfficient tuning (LoRA)
Lightweight methods that adjust a small part of the model, cutting training cost.
Discuss thisPreference tuning
Training on ratings from your experts so the model learns which answers they prefer.
Discuss thisSmaller, faster models
Training compact models that run cheaply on your own servers for one specific job.
Discuss thisEvaluation and benchmarking
Test sets built from your work to measure accuracy, style and safety objectively.
Discuss thisPrivate deployment
Hosting the tuned model in your cloud or data centre, with monitoring and updates.
Discuss thisAnna
Training or tuning an AI model needs large amounts of well-labelled examples, and labelling them at scale needs the right tool.
Selected work · AnnaAnimated illustration
AI development infrastructure: multi-format annotation platform
What we built
- A labelling tool for images, video, audio and text
- Custom labelling workflows and review comments
- AI-assisted labelling that speeds up the work
What it changed
Used to prepare the training data for computer vision projects such as satellite and aerial image analysis.
Delivered by Hoki and its engineering partners. Product and client names belong to their owners.
Discuss a similar projectThe technology we build with
We pick tools around your existing systems, your team and how the product will be run.
Base models
LlamaMistralQwenGemmaOpenAI fine-tuning API
Training
PyTorchHugging Face TransformersPEFTTRLAxolotlUnsloth
Evaluation and serving
Weights & BiasesMLflowvLLMOllamaAWS SageMakerDocker
Where this applies
The same engineering, shaped by the rules and realities of each sector.
Financial services
Models that write reports and classify documents using your exact terminology.
Healthcare
Models that structure clinical notes and use correct medical codes and terms.
Manufacturing
Models that understand part numbers, fault codes and your engineering shorthand.
Media and broadcast
Models that write headlines, captions and metadata in your editorial style.
Logistics and supply chain
Models that read trade documents and extract fields in your required formats.
Public sector and education
Models that draft in official formats and handle local language and terms.
Frequently asked questions
Do we actually need fine-tuning?
Often not. Clear instructions or letting a model search your documents solves many problems more cheaply. We test these options first and recommend fine-tuning only when it gives a clear, measurable gain.
How much data do we need?
It depends on the task. Narrow, consistent tasks can need fewer examples than broad ones. Quality matters more than volume, so we help you select and clean the right examples.
Who owns the tuned model?
Ownership is agreed in the contract. Typically the client owns the training data and the tuned model weights, subject to the licence terms of the base model we start from.
How is our training data protected?
We agree handling rules before work starts. Training can run in your own cloud account, personal details can be removed first, and data is deleted from our systems when the project ends if you wish.
Another question? Email hello@hoki.co
Check whether fine-tuning suits your task
Share a few examples of the output you need. We will tell you which approach is likely to work best.
Start a project
Tell us what you want to build.
Share a few details about your LLM fine-tuning project. We will reply with questions and a suggested next step.
- What problem you want solved
- What you already have in place
- Any timing or budget you are working to
Start a project
Tell us what you want to build
A few details help us bring the right people to the first call.
Thank you.
Your enquiry is with our team. Here is what happens next.
1We read your enquiry and bring in the people best placed to help.
2We reply by email to arrange a call at a time that suits you.
3On the call, we agree a clear next step, with scope and cost.