Skip to content
AvraAvra
Esc
↑↓navigate↵open⌘Jpreview
On this page

Adaptation and fine-tuning

How a copy of Avra RFM learns from your database or your task, when that pays, and why the copy stays exclusive to your workspace.

In-context prediction is the default: the shared model reads your history when a prediction is made, and no weights change. Where it pays, a copy of the RFM learns from your data instead, in one of two ways.

Mode What happens Effect on the model
In-context prediction Your history is read as context when a prediction is made No weights change
Adaptation Self-supervised learning on your database A copy exclusive to your workspace
Fine-tuning Learning one particular task A copy exclusive to your workspace
Your workspaceAvra RFMshared, pre-trainedYour historyread as contextAdapted copylearned your databaseFine-tuned copylearned one taskcopy
The shared model reads your history as context without changing. An adapted or fine-tuned copy is made inside your workspace, and it’s never merged back into the shared model.

Adaptation

The copy learns from your database itself, without labels: how your tables connect and what their values look like. Every question you then ask over that database starts from a model that already knows it, so adaptation helps most when you’ll ask many questions of the same database.

Fine-tuning

The copy learns one task from your labels: the outcome you care about most, at the horizon you need. Fine-tuning helps most when one prediction carries a decision that matters enough to justify its own model. It also opens up other kinds of targets, such as one category among several. See What you can predict.

When to use them

Start in context. It needs no training run, so you learn whether a question is worth pursuing before you train anything. Adapt or fine-tune when an evaluation on your history shows the copy improves on in-context results by enough to justify maintaining it, because a copy is a model of its own to evaluate, release, and keep current.

An adapted or fine-tuned copy is evaluated and released like any other version. See Evaluation and Model versioning.

Your copy

An adapted or fine-tuned copy is exclusive to your workspace and never serves another customer. It’s a one-way fork: it’s never merged back into the shared model, so nothing about your business reaches anyone else’s model.

The copy is derived from Avra’s model and licensed for use in your workspace, not delivered: its weights are never handed over as usable files. It’s stored where you run. In Avra’s cloud, that’s inside your isolated workspace. In your AWS account, the training job that produced it writes it, encrypted, to your S3.

How to run them

In Avra’s cloud, Avra runs adaptation and fine-tuning in your workspace through Avra Core. In your AWS account, they’re training jobs on Amazon SageMaker, and inference jobs then use the copy they produced. See Avra RFM on AWS SageMaker.

Was this page helpful?