Store fine-tuned model output

Keep the model from a fine-tuning job in a Model Repository. You can download it, or start a new job from one of its checkpoints.

Before you begin

• A fine-tuning job that has made a model.
• Permission to read model repositories in the project. A Member gets it through a policy that allows reading for the Inference service.

Find the job’s repository

On the fine-tuning job’s Models tab, select a model’s name, then select Continue in the Redirect dialog. You see only the repository the job made; to see all repositories, select Refresh.

After a job finishes, its TensorBoard log files are in this folder of the job’s repository:

tensorboard_logs/

The job’s page also has three setup tabs for downloading the model: Using SDK, Using TIR Notebook, and Using CLI. See Use a repository from an instance or notebook.

Continue training from a checkpoint

1. In the fine-tuning job’s form, open Model Repository Type. It offers a base model or “Continue training model from previous checkpoint”.
2. Choose “Continue training model from previous checkpoint”. Under Select Model Checkpoint, you can choose only a repository that a fine-tuning job made.
3. Choose a repository, then one of its checkpoint folders, and select Submit.

Serve the fine-tuned model

The repository a job made is a repository of your project, like any other. The Model Repository list on the endpoint form shows it, so you can link it to a Model Endpoint. See Use a repository from a Model Endpoint.

Use the repository for other models

You can also store base models for fine-tuning in a repository, and download models to instances for training. See Upload model files.