A logistics company wants a model to produce delivery-route plans.

How it works

Foundry's fine-tuning overview matches each managed fine-tuning method to the training signal you have:

• SFT: learns from prompts and desired responses. Fits when you can demonstrate the behavior.
• DPO: learns from preferred and rejected responses. Fits when quality depends on preferences such as tone.
• RFT: learns from prompts plus a grader that scores responses. Fits when many solutions are possible and you can reliably score them.

The route-planning case matches the RFT description word for word. Many plans can be valid, so there's no single answer to demonstrate, but existing code scores any plan for cost and rule compliance. That code becomes the grader. The model learns to produce plans that score well.

On the approach: managed fine-tuning runs a training job from your data and settings. The overview recommends starting there if you're new to model customization, because its workflows are predefined. Interactive training (preview) gives Python control over losses, rewards and rollouts. It's aimed at experienced ML practitioners and requires access approval.

The other options, and when you'd use them

• SFT with one sample plan per prompt: SFT learns from prompts and desired responses, for cases where you can demonstrate the behavior. The team can't write one ideal answer per request.
• Interactive training with a custom loop: It's for experienced ML practitioners building custom training workflows, and it needs access approval. The team is new to customization.
• DPO with preferred and rejected pairs: DPO fits when quality depends on preferences such as tone or completeness. Here quality is a computable score.

Exam tip

Gradeable but open-ended output = RFT with a grader. Demonstrations = SFT. Preference pairs = DPO. New to customization = managed, not interactive.

References

• Fine-tuning in Microsoft Foundry