Pick a goal — the pipeline follows
Three canonical specialization recipes. Each one composes what you just learned — PEFT, preference optimization, data strategies — into a concrete shippable pipeline. Pick your goal, click through the steps.
Why the hyperparameters differ across recipes
Notice the three recipes don't share a default. Tool-calling uses LoRA r=32, lr=2e-4, 3 epochs. Domain uses r=32 (or r=64 with CPT). Personal assistant uses r=16, low LR, 1–2 epochs. These aren't arbitrary — each number is a bet about how much the base model needs to change.
These aren't just playbooks — they are what real teams actually do
Every recipe here is the composition of techniques already covered. The tool-calling recipe is what the xLAM and Phi-4-mini teams actually shipped. The domain expert recipe is what medical/legal/financial teams do when fine-tuning for regulated deployments. The personal assistant recipe is what I'd build on a Mac tomorrow afternoon.
Act VII packages the resulting model for travel; Acts VIII–IX serve it. You're now three Acts from owning a custom specialist end-to-end.