• 1. Where Fine-Tuning Sits in the Stack

2. The Decision Tree — When to Fine-Tune

3. The Methods — Full FT, PEFT, LoRA, QLoRA

• 4. The Data Side — SFT Format and Volume

5. Preference Tuning — RLHF and DPO

• 6. Vertex AI Tuning — Concretely

7. The Customer Mismatch Patterns ⭐ (FDE gold)

• 8. Evaluating Fine-Tuned Models
• 9. Pitfalls Checklist

10. The "drop in every fine-tuning answer" checklist

When fine-tuning comes up in an RRK answer, touch:

1. Did we exhaust cheaper interventions first? (prompt eng, few-shot, RAG)
2. What's the right method? (LoRA for most cases; QLoRA if budget-constrained; DPO if it's an alignment problem)
3. What's the data story? (volume, quality, diversity, held-out split)
4. What's the eval story? (baseline before, target metric, general-capability regression check, safety check)
5. What's the serving story? (adapter swapping, endpoint management, rollback)
6. What's the maintenance story? (when do we retune? what triggers it?)