Syllabus Lesson 217 of 246 · Module 20: Adapting a trained model cheaply
Module 20: Adapting a trained model cheaply

Gate: a new task on a budget (AI off)

Why it matters Real adaptation work has a target, a budget and something you must not break. All three, at once, alone.

Gate rules: AI tutor off. Adapt the recorded GPT to Python code with at most 60 training steps of 8 windows each. Targets on a fresh run (new seed):

  • held-out code loss at most 2.43 (the base scores 3.25);
  • held-out English loss at most 2.00 (the base scores 1.70);

Set plan: method "full" or "lora", the rank (for LoRA, 1 to 64), the steps and the learning rate. The check trains your plan once on a new seed.

Words and symbols in this lesson
LoRA
low-rank adaptation: a frozen weight matrix plus two thin trainable ones, x W + (alpha / r) x A B
rank (r)
how thin LoRA's two matrices are: (d_in + d_out) x r trainable weights per matrix
Your turn

Fill plan.

AI off: no tutor, hints or worked solution on this one. Your own thinking only.

Spotted a problem in this lesson? Report it · Ask about this lesson

This exercise opens soon

Finish the lesson before it and this one unlocks. The whole course is free, no account needed, and you can read this lesson right now.