Model Training

Every time you add new data to an existing dataset or a newly created dataset, we schedule a training that will be executed within 24 hours.

The training is scheduled based on these conditions:

• No training is currently scheduled or running.
• New examples are added to a dataset.

New training is scheduled:

• Experiment phase - For the first hour (<60m)

  • immediately (+1 minute) after appending a sample

• Production phase - After the first hour (>60m)

  • training triggers every 24 hours by adding a sample

Only one training runs at a time. While a training is scheduled or running, adding more examples does not schedule another one. If new examples arrive while a training is running, a follow-up training is scheduled roughly a minute after that training finishes.

The 24 hour cadence above applies to production. On staging, trainings are scheduled far more frequently - within minutes - so that you can iterate while integrating.