liveBook · Manning

Fine-tuning

As explained in AI Governance

Fine-tuning is the additional training of a pretrained model on task-specific or organization-specific data to adapt its behavior or capabilities.

Fine-tuning adapts a pretrained model to a domain using tailored data. It can also weaken safety behaviors, including when the data is benign, so domain performance alone is not enough to establish deployment readiness. section 4.4.1

For organization-specific fine-tuning, pair customization with safeguards:

  • Validate and clean training data; assess risks of bias, poisoning, safety issues, and privacy violations. section 2.4.3
  • Retest safety behavior after fine-tuning, even when the dataset is benign. section 2.4.3

Safety can erode during adaptation

Reported tests found that ten adversarial examples weakened GPT-3.5 Turbo’s guardrails; fine-tuning Llama-2 variants on benign medical, financial, or legal data also weakened safety behaviors. section 4.4.1

How AI Governance covers it

How deep it goes
Connects fine-tuning to safety degradation and poisoning risks, but does not detail training procedures.
Examples
Cites adversarial examples against GPT-3.5 Turbo and benign medical, financial, and legal datasets used with Llama-2 variants.
What you can do after
Operate/configure it with validated, cleaned data and safety retesting after fine-tuning.
How it connects
Links fine-tuning to model weights, safety alignment, jailbreaks, and data poisoning.

Also explained in

Compare how 52 books explain fine-tuning →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.