ML Training Paths — Which One Applies

"Train," "fine-tune," and "personalize" get used interchangeably, but on Apple platforms they map to six distinct toolchains — each with a different output file, runtime, and maintenance contract. Picking the wrong one routinely wastes weeks, most often by building an MLX or .mlpackage pipeline whose output turns out not to be loadable where you need it. This page helps you pick the right path before you build.

When to Use

Use this when:

• You are about to start a training, fine-tuning, or personalization pipeline and want to confirm its output will load on your target.
• You are unsure how Foundation Models adapters, Create ML, MLUpdateTask, coremltools conversion, and MLX relate.
• A trained artifact will not load and you suspect a format or toolchain mismatch.

Example Prompts

• "What's the difference between training a Foundation Models adapter, fine-tuning with MLX, and personalizing with MLUpdateTask?"
• "I fine-tuned a model with MLX — how do I load it into Foundation Models?"
• "Can I personalize my .mlpackage on device?"
• "Which training path ships to iPhone, and which are Mac-only?"

The Six Paths

Path	What it produces	Use when
FM custom adapter	.fmadapter for Apple's on-device LLM	App-specific LLM behavior, after simpler approaches fail
Core ML personalization (MLUpdateTask)	Updated .mlmodelc	Per-user tuning of an existing NN-spec (not .mlpackage) model, on device
Create ML	A new .mlmodel	Training a task model from scratch
coremltools convert	.mlpackage	Bringing an already-trained PyTorch/TF model to Apple platforms
MLX LM	adapters.safetensors (Mac-only, not an iOS path)	Research and on-Mac experimentation
Server LLM fine-tune	A cloud artifact	Vendor cloud-model customization

The Traps That Cost Weeks

• MLX output is not a Foundation Models adapter – mlx_lm.lora emits adapters.safetensors; the on-device LLM loads only .fmadapter. Different toolchains — MLX cannot feed SystemLanguageModel(adapter:).
• MLUpdateTask is NN-spec only – it does not apply to ML Program (.mlpackage) models, which is what modern conversion produces. Decide the format before building.
• FM adapters pin to one base-model version – they must be retrained and re-shipped each OS minor that changes the base model; Core ML, Create ML, and MLX models carry no such pin.

The skill walks each path end to end, with a decision tree and the full trap list.

Related

• CoreML – the toolchain behind the Create ML, MLUpdateTask, and coremltools-conversion paths
• Foundation Models – start here for on-device LLM work; adapter training is the last resort
• Foundation Models Adapters – the how-to for the .fmadapter path
• iOS ML deployment overview – where each artifact actually runs on device