Fine-tuning vs Prompting

What you'll learn

• Choose between prompting, retrieval and fine-tuning using an evidence-based decision ladder.
• Explain how SFT, learning rate, LoRA, preference tuning and distillation change a model.
• Prepare, validate and split training data without noise or leakage.
• Spot overfitting, catastrophic forgetting and safety regressions, and apply fixes such as early stopping and replay.
• Compare a tuned model with a strong prompted baseline and compute the break-even volume.
• Plan serving, versioning, monitoring and retirement for a tuned model.

Syllabus

Choosing Between Prompting, Retrieval and Fine-Tuning

1. Three Levers and What Each One Changes
2. The Decision Ladder: Cheapest Effective Step First
3. When Fine-Tuning Is the Right Tool
4. The Economics: Break-Even Volume

How Far Prompting Goes

1. Few-Shot Prompting: In-Context Learning
2. Learning Curves: When Training Overtakes Examples in the Prompt
3. The Cost of Long Prompts and Prompt Caching

How Fine-Tuning Works

1. Supervised Fine-Tuning (SFT) in Plain Terms
2. Learning Rate: The Most Sensitive Knob
3. LoRA and Parameter-Efficient Fine-Tuning
4. Overfitting, Epochs and Early Stopping
5. Preference Tuning: RLHF and DPO
6. Distillation: Teaching a Small Model From a Large One

Training Data: The Real Work

1. Label Quality: Noise Costs Accuracy
2. Dataset Format and Validation (JSONL)
3. Train, Validation and Test Splits Without Leakage

Risks: Forgetting, Facts and Safety

1. Catastrophic Forgetting and Replay
2. Fine-Tuning Is a Poor Way to Add Facts
3. Safety, Privacy and Licensing of Training Data

Evaluating and Operating a Tuned Model

1. Always Compare Against a Strong Prompted Baseline
2. Choosing Metrics for the Task
3. Serving Tuned Models: Merged Weights and Adapters

Putting It Into Practice

1. Case Study: Support Ticket Routing
2. Hosted Fine-Tuning versus Open-Source Training
3. Versioning, Drift and Retirement

Pitfalls and a Final Checklist

1. Common Pitfalls
2. A Decision Checklist