Training a Base Model
Finetuning adapts a base model that has some, but not all, of the capabilities needed so that it performs well for a specific task.
A Model That Knows Part of the Job
Imagine a model that already knows something useful but not everything you need. It can handle part of the work, yet it falls short on the particular job you care about. Finetuning is the process that closes that gap. It begins with a base model that has some of the capabilities the job requires, and it adapts that model so it performs well enough on one specific task.
What do you think happens?
A base model has some of the capabilities a specific task requires. Does it already have all of them?
Reveal answer
No, it has some of the required capabilities but not all of them
This gap is exactly why finetuning exists. The base model is a useful starting point, but it still falls short on the particular job, so it must be adapted before it performs well for that task.
Tracing a Model Through Its Two Stages
Transfer learning, the strategy behind finetuning, runs in two stages. In the first stage, a base model is trained on tasks where plenty of data is available. In the second stage, what that model has learned is transferred to the target task, which is the task you actually care about but for which data is scarce. Follow the model's state through the diagram below: it starts as a general-purpose base model, passes through a finetuning step that uses data from the target task, and ends as a finetuned model.
The diagram shows the key idea of the whole article. The base model is not thrown away. It is the starting point for the finetuning step, which is why the earlier stage matters so much to the final result.
Finetuning Inside Transfer Learning
Finetuning is best understood as one way of doing transfer learning. Transfer learning is the broader idea of taking knowledge gained from one task and using it to speed up learning on a new, related task. The concept was first introduced by Bozinovski and Fulgosi in 1976. Finetuning is therefore a specific route within that larger idea, not a separate subject standing next to it.
Google's multilingual translation system shows transferred knowledge enabling translation between a language pair with no direct training examples. The system learned from many language pairs, and that learned knowledge carried over to a pair it had never been trained on directly.
What Carries Over From the Base Model
Transfer learning helps most when target-task data is limited or expensive, because a base model trained on data-rich tasks can pass its knowledge along. Finetuning does not start from nothing. The capabilities the base model already has are the foundation, and adaptation builds the missing pieces on top of that foundation for the specific task.
Adapting a General Model to a Support Desk
A base model can read and write general language but has not been trained on a particular company's support tickets. Which part of its ability carries over, and which part must be adapted?
- 1Identify the capability the base model already has
The base model already handles general language. That part is the knowledge carried over from its earlier training, so the team does not need to teach it language from scratch.
- 2Identify the required capability that is missing
The support desk needs answers in the company's own format and terminology. The base model does not have all of this, so it is one of the gaps that finetuning must close.
- 3Adapt on the target task
Finetuning uses the target task data, which is the company's ticket examples, to adapt the base model so it performs well on the support desk job.
The general language ability carries over unchanged in purpose, while the task-specific behavior is added through finetuning. The result is a finetuned model that serves the support desk, built from a base model that already had part of what the job required.
Base Model Versus Finetuned Model
The base model and the finetuned model are closely related, but they are prepared for different levels of usefulness. The base model has some capabilities, though not all of those the task requires. The finetuned model is the result of adapting that starting point so it performs well for the specific task. The side-by-side view below contrasts the two states.
| Aspect | Base model | Finetuned model |
|---|---|---|
| Starting point | Trained on data-rich tasks | Begins as a base model |
| Capabilities | Some of those the task requires | Adapted so the task is handled well |
| Prepared for | Partial usefulness across the work | One specific target task |
| Role | Source of transferred knowledge | Result of adapting that knowledge |
Base model and finetuned model compared by the task each is prepared to perform
Mistakes About Finetuning and Transfer Learning
Common mistake Believing finetuning teaches a model entirely new knowledge
Why it failsFinetuning starts from a base model that already has some of the required capabilities. Adaptation closes the gap; it does not replace what the model already knows.
Do this insteadThink of finetuning as adapting existing capabilities toward a specific task.
Common mistake Treating a base model as finished for every task
Why it failsA base model has some capabilities, though not all of those the task requires.
Do this insteadCheck whether the base model covers the specific task before relying on it directly.
Common mistake Treating transfer learning and finetuning as separate subjects
Why it failsFinetuning is one way of doing transfer learning. Transfer learning is the broader umbrella.
Do this insteadPlace finetuning inside transfer learning as one specific method.
Common mistake Assuming transfer learning is only needed when data is plentiful
Diagnosing a Model's Gap
A team has a base model trained on data-rich tasks. It handles general language well but cannot yet answer questions about their product catalog in the required format. Using the ideas from this article, explain which capability the base model already has, which required capability is missing, and what finetuning should do about the gap.
Hints
- Ask which part of the job the base model can already handle
- Ask what the target task requires that the base model does not yet have
- Remember that finetuning adapts the base model rather than starting over
Key Takeaways
- 1A base model has some, but not all, of the capabilities a specific task requires
- 2Finetuning adapts the base model so it performs well on one specific task
- 3Finetuning is one method inside transfer learning, which uses knowledge from one task to speed up learning on a new, related task
- 4Transfer learning is most useful when target-task data is limited or expensive
At a glance
- A base model has some, but not all, of the capabilities a specific task requires
- Finetuning adapts the base model so it performs well on one specific task
- Finetuning is one method inside transfer learning, which uses knowledge from one task to speed up learning on a new, related task
- Transfer learning is most useful when target-task data is limited or expensive