Skip to main content
  1. Prepare. Import your catalog and events using the data format.
  2. Choose. Select a base model and compatible tokenizer with the team.
  3. Train. Create an adaptation job for next-item prediction.
  4. Evaluate. Compare against popularity, a task baseline, and the unadapted model on a temporal holdout.
  5. Activate. Receive a new model ID after the evaluation is reviewed. The base version stays unchanged.
The result includes the dataset snapshot, split dates, tokenizer, model version, and report. A completed training job returns a candidate model; it does not silently replace a live deployment.

​
What changes

Sending a history conditions one request. Adaptation changes model weights using an approved dataset. Neither operation promises better results on every task. Adaptation is a managed beta workflow. Timing and supported base models are confirmed at onboarding. Training from scratch is identified separately from adapting a pretrained model.