Custom models
Your task needs
its own model.
You keep the weights.
We train a model for one job: sort a queue, pull fields, score a file, or answer in your wording. You bring the examples. The weights leave with you.
A public chatbot is a generalist. A custom model is trained on your labelled examples for one outcome. Most language work starts from an open-weight base and adapts. A narrow task can be a smaller model when a full language model is the wrong tool. We will say so before any training spend.
This is model creation for a Kenyan team, with the corpus staying in Kenya when it must, or on a cloud GPU when counsel says it can travel. Deeper language adapters live on fine-tuning. Putting the finished file behind an API is model hosting.
How a model gets built
Define the job, then train, then score.
$ corvux model train --task claims-sort --split held-out BASE open-weight or a smaller task model DATA your labelled rows · PII redacted RUN cloud GPU or Kenya hardware, named on the quote EVAL held-out set you approve ✓ weights and recipe assigned to you
What we train
One job, one model, a person still in the loop.
Classifiers and extractors
Route a ticket, pull a field, flag a mismatch. A small model when a script or a prompt already fails the bar you set.
Language models
Start from Llama, Qwen, Gemma, Mistral, or another open-weight base, then fine-tune on your wording.
Assistive vision
Image workflows where a clinician or operator still decides. We ship assistive software, and we leave diagnostic-device claims off the page.
The file is yours
Weights, config, and the training recipe go to your repo or tenant. Corvux keeps no pooled model trained on your files.
Related: eval and red team before go-live, model routing when a small model should take the easy prompts, and how we quote once we have seen the corpus.
FAQ
Before you send a dataset.
How many examples do we need?
Hundreds of clean labelled rows can train a narrow task. A language adapter wants more. We read the set before we promise a gain.
Who owns the model?
You do. The contract assigns the weights and the training recipe to you. Corvux keeps no pooled model trained on your files.
Where does training run?
On a cloud GPU when the corpus can travel. On Kenya hardware when it must stay. The quote names the place before the first run.
Is this the same as fine-tuning?
Fine-tuning adapts an open language model to your wording. This page also covers smaller task models, classifiers and extractors, when a full language model is the wrong tool.
Bring the task and a sample
A dozen real inputs and the output you want is enough for a first read. We will say whether to train, fine-tune, or skip the model.
Scope a custom model