Training a model from scratch or adapting an existing one

What should belong to the company

In a hypothetical company, an app helps employees prepare descriptions of finished projects. The director wants to be able to develop it independently of any one supplier. That goal needs access to the app's code, the documentation and knowledge of its dependencies. It does not yet settle how the language model was created.

If the problem is where it runs, the team can assess a model that can be run independently under the terms of its licence. If the answers have to follow the company's way of describing a project, start with instructions and checked examples. When current facts about the project are needed, the app has to receive the right information.

Each of these options gives a different scope of control. Access to the model's files alone does not mean the company has a team able to maintain it, assess new versions and fix problems after the environment changes. These duties should have an owner already in the option of cooperation you are considering.

What training a model from scratch changes

Fine-tuning takes a model that has already been trained and trains it further on examples related to the task. Training a language model from scratch also covers building its basic skills from the training material. It is a much wider undertaking than adjusting the style of project descriptions.

The Hugging Face course explains that training from the start can make sense with a large body of material very different from the data of existing models, and needs far more compute than fine-tuning. This is a general distinction for language models; the choice for a company needs an assessment of its task.

The decision also has to include the work after the first launch. Who will find the model's weaknesses, assess changes and prepare the next versions? Does the company want to keep such a capability, or does it mainly need a working app for a specific task? The answers shape the scope of the order much longer than the initial training does.

What should remain after assessing the options?

Ask for a record of the task, the base model, the data version and the comparison criteria. In a fictional example, the problem is missing information in project descriptions. If the missing facts never reached the input, training does not fix how the material is passed on. Check a fuller source first.

The report should also keep the results that argue against training. If adaptation is justified, settle who supplies the correct examples and who keeps the assessment going after the process changes. Running ready weights on their own follows a separate path, described in the first trial of a local model.

When adapting an existing model is worth a feasibility assessment

For project descriptions, the starting point can be a comparison of models with the same source material and instructions. Employees check whether the description keeps the facts and how much editing it needs. If recurring problems remain, adaptation can be assessed. If the limits concern access to data or the connection to the project system, the app needs its own scope.

Training from scratch needs a justification that these simpler options do not meet. In the proposed assessment, we name such a requirement directly and describe what resources the company would need to meet it. Wanting to have “our own AI” is too general to choose a scope of work.

Start the first conversation with what the company wants control over and what current solutions do not give it. We can then propose an assessment of the right options, without settling on training in advance. We price a private model after a call about the data and the hardware, and an app on it starts from €3,000 net (pricing).

What a model in the company costs and when it pays back

Where to start

In the free process review we choose one process and 20 to 100 cases the team has already handled. We compare the simpler change with the adapted model on cases the model has not seen, and within 2 working days you get a written summary with a price. We adapt models as part of private LLMs, and build the app as an AI app.