[ Fine-tuning ]
How a training project runs
How a training project with Kontinent runs: from your documents to a model of your own, measured before and after.
Besides the API, Kontinent trains models for one specific task of yours: extracting fields from a document, classifying it, or making a decision whose correct answer can be written down and checked. This is a project we run together with you, not a self-service feature. There is no training API and no training job in the console yet: the Training, Dedicated and Datasmith pages in the console are previews of what is coming.
To start a project, book a call.
Terms used on this page: the small model trained for you is the student, the large model that helps it learn is the teacher. The correct answers you provide for a sample are gold labels. The measurement before training is the baseline, the one after it the eval. An adapter is the file that holds what the model learned.
What you get
- The adapter, in the standard safetensors format. It runs with common open tooling (transformers, vLLM, TGI) on the public base model. You do not need Kontinent to run it, and it belongs to you.
- Two measured numbers: how well the base model did your task before training and how well your model does it after, on the same documents, with confidence intervals and a statement whether the difference holds statistically.
- A model card: base model, license, which data was used, results before and after, known limits, how to run it.
- An evaluation log: every test document with the answer before training, the answer after, and your correct answer, deviations marked.
- Checksums over everything, so you can prove the package is unchanged.
How a project runs
| Phase | What happens | Your effort |
|---|---|---|
| 1 Define the task | Together we define what the model should return for every document. | One session of about 90 minutes with a domain expert |
| 2 Deliver documents | You hand over 500 to 1,500 documents without answers. | Export from your system, clearing the data transfer |
| 3 Label a test set | You give the correct answer for 40 to 90 documents. | About two minutes per document |
| 4 Measure before training | Baseline, the teacher's ceiling, diagnosis. | None, you receive the numbers |
| 5 Choose the method | We recommend a method and quote the cost; you decide. | One decision |
| 6 Train, measure, hand over | Training run, eval, package. | Acceptance of the result |
After phase 4 there is a point where we stop before you pay for compute: if even the teacher does not beat what your current system already achieves, training a student on it will not help either, and we tell you so with the numbers.