Core ML Training and Personalization

Two different ways to train a Core ML model that developers often mix up: Create ML trains a new model from your data, and MLUpdateTask adapts a model you've already shipped to each user's data, on their device.

When to Use

Use this skill when you're:

• Training an image, sound, text, or tabular model without writing a training loop
• Choosing between the Create ML app and the CreateML framework
• Making each user's copy of a model adapt to their own data, on-device
• Planning on-device personalization for a model you converted to .mlpackage, which MLUpdateTask can't update

Example Prompts

Questions you can ask your assistant that will draw from this skill:

• "How do I train an image classifier from a folder of labeled photos?"
• "Can I train a Create ML model on iPhone, or only on a Mac?"
• "How do I personalize my model on the user's device with their data?"
• "Why can't I use MLUpdateTask with my converted PyTorch model?"
• "Training freezes my app's UI. How do I run it without blocking?"

What This Skill Provides

• Create ML – The app and the CreateML framework, the model types each supports, and which ones train on iOS, iPadOS, and visionOS as well as macOS
• Programmatic training – Data sources, exporting a .mlmodel, and the async training session that keeps a long run off the main thread
• On-device personalization – How MLUpdateTask retrains a deployed model's last layers on a user's data and saves the result
• The NN-spec limitation – Why MLUpdateTask only updates NeuralNetwork-format models, not the ML Program (.mlpackage) models modern conversion produces, and what that means before you build a pipeline
• A boundary table – How Create ML, MLUpdateTask, Foundation Models adapters, and model conversion differ

Documentation Scope

This page documents the coreml-training skill (in the axiom-ai suite), which your assistant loads when you train or personalize a Core ML model. Training a custom adapter for Apple's built-in language model is a different toolchain, covered by foundation-models-adapters.

Related

• ios-ml – The hub that routes Core ML work to conversion, compression, training, and deployment
• training-paths – Choosing among Create ML, MLUpdateTask, adapters, and the other training toolchains before you build
• coreml-conversion – Converting a model you trained elsewhere instead of training it with Create ML
• coreml-compression – Shrinking a trained model before you ship it

Resources

WWDC: 2018-703, 2019-430, 2019-704, 2021-10037, 2024-10183

Docs: /createml, /createml/mlimageclassifier, /coreml/mlupdatetask, /coreml/mlupdatecontext