Table of contents
Fine-tuning lets you take an existing model and train it with your own data to create a new model that is better suited to a specific task. Whenever you create a new fine-tune, you must specify a destination
model. This is a model that you create, and it will be updated with the results of your fine-tuning process. The destination model can be a new model, or an existing model that you've already fine-tuned.
This guide covers how to create and specify destination models for fine-tuning.
Every trainable model has a "Train" tab with a form that lets you create a new fine-tune of that model. For example, check out the training pages for models like Flux or LLaVA:
The training form asks you to choose a destination model. You can select an existing model as a destination, or create a new model.
If you want to create a destination model without training it right away, you can create a model manually from the web at replicate.com/create.
You can use the HTTP API to create a new model programmatically whenever you need a new destination for a training.
Here's an example using cURL:
You can also use Replicate's JavaScript and Python clients to create models with the API.
You don't have to create a new model every time you fine-tune. You can instead use the same destination model multiple times, and the resulting fine-tunes will be added to the destination model as different "versions". You can think of the destination model as a collection of fine-tunes, in the form of model versions.
You only need to create the destination model once. When a new training is created, the resulting fine-tune is automatically pushed as a new version to the destination model.