ai-agent-book/slides/lesson-27.md

theme, title, info, author, transition, mdc, lineNumbers, monaco, aspectRatio, canvasWidth, layout, class

theme	title	info	author	transition	mdc	lineNumbers	monaco	aspectRatio	canvasWidth	layout	class
seriph	Lesson 27 — When Should You Teach with Examples—and When with Rewards?	English video course for AI Agents in Depth	Bojie Li	slide-left	true	false	false	16/9	980	cover	cover

Improve · Chapter 7 · Model Post-Training

When Should You Teach with Examples—and When with Rewards?

SFT, loss masking, distribution shift, and the form-first rule

Lesson 27 of 42 · 19 minutes · SFT; When to Choose SFT and When to Choose RL; Single-Turn Reinforcement Learning

---

layout: center class: text-center

The central question

Is the target capability a stable mapping to imitate or a strategy that must survive new situations?

---

Why this problem matters

SFT

Dense token-level supervision; stable and sample-efficient.

RL

Sparse trajectory feedback; costly but allows exploration.

Shift

Test whether the learned behavior survives changed rules and inputs.

---

Three ideas to keep in view

Loss masking

Supervise the assistant response—not the user prompt

Form first

Stabilize parsable output before optimizing strategy

Holdout shift

Change values or environments while preserving the rule

---

The book's visual model

SFT followed by RL as a two-stage training pipeline

---

Use SFT vs. Consider RL

Use SFT

• Known demonstrations
• Format/style/protocol
• Deployment matches training

Consider RL

• Outcome can be verified
• Many valid strategies
• Generalization under shift matters

SFT and RL are sequential tools—not rival ideologies.

---

SFT masks the prompt tokens

tokens = prompt_ids + response_ids
labels = [-100] * len(prompt_ids) + response_ids
loss = cross_entropy(model(tokens), labels)
loss.backward()
# only the demonstrated response is supervised

---

Test the claim

7-4 evidence2 min

Audit retained VLM pre-training and SFT evidence

Observe: Hashed outputs, blind judgments, matched configurations, and negative results

7-5 evidence2 min

Audit continued-pretraining trade-offs

Observe: New-language gain, retained English ability, and persistent factual errors

Demo budget: 4 minutes · one contiguous terminal block

---

class: course-terminal

Live demo

Switching to the terminal

$ python chapter7/MiniMind-pretrain/validation/validate_vlm_evidence.py

$ python chapter7/continued-pretraining/validation/validate_evidence.py

Run the command(s), narrate decisions, and point to the observation—not just the output.

---

What the evidence supports

Finding 1

SFT efficiently learns explicit protocols represented in examples.

Finding 2

RL is justified when a verifier can reward strategies beyond one reference answer.

Finding 3

Training gains must be tested beside retention and distribution-shift failures.

---

layout: center

Where the claim stops

Boundary condition

The slogan 'SFT memorizes, RL generalizes' is a tendency under controlled conditions—not a guarantee for every model and task.

---

layout: center

Engineering takeaway

Design rule

Use SFT until output is stable; add RL only when exploration and verified generalization justify its cost.

---

Continue the experiment

Experiment 7-3: MiniMind language training chapter7/MiniMind-pretrain/ Experiment 7-11: SFT vs. RL reproduction chapter7/SFTvsRL/ Experiment 7-6: Sesame speech SFT chapter7/sesame/ Experiment 7-6: Orpheus speech SFT chapter7/orpheus/ Experiment 7-7: multilingual reasoning chapter7/MultilingualReasoning/

---

layout: center class: text-center

Pause and apply

Your turn

What deployment change would reveal that your fine-tuned model learned an example instead of a rule?

---

layout: center class: text-center

Next · Lesson 28

Translate preferences and outcomes into optimization signals without losing the base policy.

→