AIM Fine-tuning

AIM Fine-tuning makes fine-tuning different AIMs easy by providing a common interface for supported training engines. This makes it possible to train different models using different training methods without learning the implementation details of each engine.

For a step-by-step example of running a training job, see the Quickstart.

Fine-tuning adapts an already trained AI model to suit a new purpose or learn a new capability. It is an inherently experimental, iterative process involving three key elements:

• The model, which should have the capabilities necessary to learn what we want it to learn.

• The data, which should contain clear examples of the behaviour we want to teach the model.

• The learning process parameters, or hyperparameters, that control how the model learns from the data. Fine-tuning involves finding suitable choices for each of these three elements.

Supported training methods, engines, and models

Training methods

• Supervised fine-tuning (SFT)

Training engines

• aimfttk: AIM Fine-tuning Toolkit engine, available in the amdenterpriseai/aimft-aimfttk Docker image

Models

AIM Fine-tuning provides predefined recipes that specify training parameters for selected models and training engines. See the Supported models page for the supported model, training method, engine, and recipe combinations.