Pradra StudioDocs
Sign in

Training

This is where a capability learns. Pradra Studio runs the training for you and keeps a complete, tamper-proof record of how each model came to be.

Starting a run

To train, you point a capability at a dataset version and a feature set version, choose an algorithm, and start the job. The platform handles the compute; you watch the progress.

  1. Go to Training in the sidebar (or start from the capability's detail page).
  2. Pick the capability, then the dataset version and feature set version it should learn from.
  3. Choose an algorithm. A sensible default is provided — you don't need to be a specialist to get a first model — and hyperparameters are adjustable when you want control.
  4. Submit. The run is queued, then trains; the list polls live, so you see the status change without refreshing.

What happens during a run

  • The platform materializes the training data from your exact dataset + feature set versions (derived features included, with a frozen clock so the run is reproducible).
  • The model trains on one part of the data and is scored on a held-out part it never saw.
  • Everything about the run — parameters, metrics, logs, artifacts, the exact input versions — is recorded as an experiment.
  • The run produces one candidate model, which enters evaluation.

When a run fails

A failed run is a fact, not a catastrophe: the error is recorded on the run, nothing reaches production, and nothing you had in production is affected. Common causes are data-shaped — a target column with too few examples, or a feature column that turned out mostly empty — and the dataset detail page's column statistics are the fastest way to spot them.

You can also cancel a queued or running job from the Training page.

Retraining is the same thing

The automatic retraining path doesn't get a special pipeline: it submits an ordinary training run with the same algorithm and settings as the current production model, on fresh data, and the candidate faces the same gate. One pipeline, no side doors.

Next