Fine-Tuning — Mostly Don't | Beyond the Prompt

Lesson 7: Fine-Tuning — Mostly Don't

Blackboard

Know the trade-offs before you commit

9 min

Concept 1

The case against fine-tuning

Fine-tuning requires labelled data, compute time, and expertise. It is expensive and slow. By the time you finish, the next base model may already beat your fine-tuned version of the old one. Prompt engineering methods are model-portable; fine-tuning is not.

Concept 2

Overfitting and lost generality

A fine-tuned model may perform well on your specific task but lose its ability to handle general questions. You trade breadth for depth — sometimes too much depth. The model becomes a specialist that cannot do anything else.

Concept 3

The Slack messages cautionary tale

An engineer fine-tuned a model on his company's Slack messages, hoping it would adopt the team's communication style. Instead of producing useful work, the model started mimicking how people talk about work — procrastinating, making excuses, and stalling. It learned the culture of conversation, not the skill of execution.

Concept 4

When fine-tuning does make sense

If your task requires repeated high-precision outputs in a specialised domain — legal language, scientific explanations, medical terminology — and the general-purpose model consistently fails despite good prompting, fine-tuning may be justified. But exhaust prompting and RAG first.

Pause and Think

If a better base model ships next month, what happens to your fine-tuned version of the old one?

Flashcards

Key Terms

Click or press Enter to flip a card. Review these before the practice question.

Additional Resources

Paper — RAG vs Fine-Tuning Tradeoffs
RAG vs fine-tuning compared head-to-head using agricultural data. Breaks out the cost, accuracy, and latency numbers — the kind of evidence this lesson's advice is based on.
Paper — Modifying Beliefs via Fine-Tuning
Anthropic alignment research showing fine-tuning can shift a model's internal beliefs in ways you didn't intend. A good cautionary read alongside the 'when fine-tuning makes sense' section.
Glossary — Fine-tuning
A short definition.
Glossary — LoRA (low-rank adaptation)
The cheap fine-tuning method you will see mentioned most.

Practice Question

An engineer fine-tunes a model on company Slack messages. The model starts mimicking how people procrastinate instead of doing the task. What went wrong?

Does not count for certificate

Choose one answer, use the hint if needed, then check your answer.

Use the hint before checking if you need a nudge.

Choose one answer, then click Check Answer.