Fixed project
AI production sprint
A fixed build: one system in four to six weeks. Private model / fine-tuning can follow it if you need a private model.
Read more →
Fine-tuning - Fine-tuning an open model on your own data. We run it for you, in your own cloud account
Private model / fine-tuning is for a software company. Two to six weeks. We run it for you, in your own cloud account: evaluate first, fine-tune an open model only if the hosted model falls short, and deliver an endpoint and a comparison report. It can start on its own or after an AI production sprint. Contact us for pricing. Self-hosting is available on request, on Amazon Web Services, Microsoft Azure, Google Cloud, GleSYS or in your own on-premises environment.
One task
Tell us where the hosted model falls short
A rule about where the data may live, a bill that grows with volume, or a narrow task you run over and over. We say whether Private model / fine-tuning should start on its own or after an AI production sprint. Contact us for pricing.
The fit - When a private model fits
Four reasons to look at Private model / fine-tuning. If none of them is true, we recommend you stay on the hosted model.
- We run it for you. We run it for you, in your own cloud account. Self-hosting is available on request, on Amazon Web Services, Microsoft Azure, Google Cloud, GleSYS or in your own on-premises environment. It runs where your data is allowed to be.
- Cost at high volume. The same narrow task, called often enough that the bill is the question. The report compares cost with the model you use now. We do not quote a saving before that comparison exists.
- A narrow, repetitive task. The examples look like each other. A fine-tune is for that kind of work. A broad assistant is a different job, and it stays on the hosted model.
- A domain the hosted model handles poorly. A domain the hosted model handles poorly: your terms, formats or language, where the evaluation shows it gets things wrong.
The decision comes before any fine-tune
We train only after an evaluation on your own examples says the hosted model is not good enough.
1. Start with an evaluation
On its own, or after an AI production sprint. The evaluation step ends with a written decision you can walk away with.
See the AI production sprint2. Read the evaluations
We run the hosted model on your own test examples (golden evaluation set). The question is whether it falls short on the task you actually have. That call is written down.
3. Fine-tune only if it falls short
If the evaluations show the hosted model is not good enough, we fine-tune an open model on your own examples. If it is good enough, we recommend you stay on it, and we do not train.
4. We run it for you
We run it for you, in your own cloud account. Self-hosting is available on request, on Amazon Web Services, Microsoft Azure, Google Cloud, GleSYS or in your own on-premises environment. You get an endpoint, and a simple report that compares it with your current model on accuracy, cost and speed.
What you leave with - A decision, and a model only if the decision says so
It can start on its own or after an AI production sprint, and it ends with a written decision. The endpoint and the report follow only when we fine-tune.
- The decision, in writing. Fine-tune, or stay on the hosted model. The evaluations are the reason.
- An endpoint. If we fine-tune, you can call the model where it is hosted. If we do not, there is no new endpoint.
- A simple comparison. If we fine-tune, a simple report sets the new model next to your current one on accuracy, cost and speed.
Why the fine-tune is a decision, not a default
A private model is the right tool for some tasks. It is the wrong default for most of them.
- We run it for you
- We run it for you, in your own cloud account. Self-hosting is available on request, on Amazon Web Services, Microsoft Azure, Google Cloud, GleSYS or in your own on-premises environment. The model stays where the data has to stay.
- Evaluations first
- We do not fine-tune because a private model sounds safer. We fine-tune when the hosted model falls short on your own test examples (golden evaluation set).
- Your examples
- The fine-tune uses the examples you give us for this model. We do not train on your data for anything else.
Where Private model / fine-tuning sits - Private model / fine-tuning can start on its own or after an AI production sprint
It can start on its own or after an AI production sprint. Care, a coding setup, a person in the team, or a platform takeover are the other shapes.
Related insights - From a hosted model to one you can trust
How we think about shipping one system and knowing whether a change is actually better.
Questions - Common questions
Short answers before you treat a private model as the plan.
Where does the model run?
We run it for you, in your own cloud account. Self-hosting is available on request, on Amazon Web Services, Microsoft Azure, Google Cloud, GleSYS or in your own on-premises environment. We do not move the examples off that environment to train somewhere else.
Is Private model / fine-tuning its own project?
Yes. It can start on its own or after an AI production sprint. On its own it starts with an evaluation; after a sprint it can use that sprint's evaluations. Either way, nothing is trained until the evaluation says so.
What does it cost?
Contact us for pricing. It can start on its own or after an AI production sprint.
Do you always fine-tune?
No. We fine-tune only when the evaluations show the hosted model falls short. Otherwise we recommend you stay on the hosted model, and we do not train.
What do we get?
A written decision either way. If we fine-tune, you also get an endpoint and a simple report comparing the new model with your current one on accuracy, cost and speed.
Do you train on our data?
Only on the examples you give us for this model, and only if the evaluations say we should fine-tune. We do not train on your data for anything else, and we do not train at all if the hosted model is good enough.
Ready to scope it?
Bring the task and the examples
We will tell you straight if the hosted model is enough, if Private model / fine-tuning should start on its own or after an AI production sprint, or if we are the wrong people.
Let's talk about what you're building
