Fine-tuning

When to change the model instead of its input, and how to prove it was worth it.

Before you fine-tunePrompting, then examples, then retrieval, then fine-tuning, in that order and for a reason.
Tuning dataA thousand clean examples beat twenty thousand dirty ones, and how to build and split the set.
LoRA and parameter-efficient tuningWhy training a small slice of the weights is enough, and the three knobs that matter.
Running a tuning jobFrom out of memory to a readable loss curve, and the settings you change first.
Did it actually improveA frozen eval set, a before and after, and the catastrophic forgetting loss never shows.
Project: a tuned modelTune a small model on one Arabic task and prove with a number that it beat prompting alone.