Use this lesson as a gate. If you can walk through every checkbox with a concrete example from your notes or lab, you are ready to move on. If not, jump back to the matching lesson instead of collecting more checkpoints.
Intuition
Module 3 had three themes:
Choose the right lever — prompt vs RAG vs fine-tune; SFT vs preferences.
Why might training loss fall while product quality falls?
When is DPO a better next step than more SFT rows?
What breaks if you serve an adapter on the wrong base revision?
How does alpha/r interact with a rank sweep?
What is the first file you look at in a fine-tune incident?
In code
A programmatic checklist you can tick in a notebook.
CHECKS = [
("lever_choice", "Explained prompt vs RAG vs FT with one real failure"),
("sft_loss", "Stated what tokens receive SFT loss"),
("peft_formula", "Explained LoRA update W + scale*BA"),
("recipe", "Wrote a starting LoRA hyperparameter set"),
("baseline", "Saved prompt-only baseline metrics"),
("anchors", "Ran anchor regression on a checkpoint"),
("serving", "Chose merged vs adapter with a reason"),
("writeup", "Produced ship/no-ship with a scorecard"),
]
def review(done: dict[str, bool]) -> None:
missing = [label for key, label in CHECKS if not done.get(key)]
if not missing:
print("Module 3 gate: PASS")
else:
print("Module 3 gate: incomplete")
for m in missing:
print(" -", m)
review({
"lever_choice": True,
"sft_loss": True,
"peft_formula": True,
"recipe": True,
"baseline": True,
"anchors": False,
"serving": True,
"writeup": False,
})
What goes wrong
Skipping the lab writeup — Knowledge stays abstract and evaporates.
Collecting adapters like stickers — No eval owners, no rollback plan.
Treating Module 3 as only LoRA trivia — Judgment about when not to train is the main skill.
Never rechecking anchors — Silent forgetting in later modules' projects.
Confusing preference tuning with fact updating — Still not a CMS.
Map lessons to skills
Skill
Lessons
Lever choice
3.1.1
Data / templates
3.1.2
SFT mechanics
3.1.3
Freeze vs full
3.1.4
Eval / forgetting
3.1.5
RLHF / DPO map
3.1.6
PEFT intuition
3.2.1
LoRA / QLoRA
3.2.2
Soft prompts
3.2.3
Recipes
3.2.4
Serving
3.2.5
Lab loop
3.3.1–3.3.3
Capstone readiness
Module 4–5 projects will assume you can say "we tried prompt+RAG, measured X, then LoRA'd Y with scorecard Z." Practice that sentence with your lab numbers. If you cannot fill X/Y/Z, revisit the lab—not the architecture diagrams.
Time-boxed review
Spend 45 minutes: 15 on fundamentals checkboxes, 15 on PEFT, 15 rewriting your lab decision in five lines. If any checkbox takes more than two minutes of hand-waving, open that lesson and add an example from your domain before moving to Module 4.
One-line summary
Finish Module 3 by verifying you can choose levers, run PEFT deliberately, and prove outcomes with baselines, anchors, serving plans, and a clear writeup.
Key terms
Review gate — Explicit checklist before advancing modules.
Lever choice — Prompt vs retrieval vs weight updates.
Scorecard discipline — Metrics that include task lift and regressions.
Artifact hygiene — Naming, metadata, and rollback for models/adapters.
Oral self-test — Short spoken explanations that expose fuzzy understanding.
Ship/no-ship — Binary offline decision before canary traffic.
Next-step backlog — Documented follow-ups after a no-ship or partial win.