Fine-tuning (and when not to) interview questions

Training a pretrained model further on example input→ideal output pairs so it imitates them.

The most common kind of fine-tuning.

Better prompting with examples

Prompting is cheaper, faster to iterate, and often enough.

You need consistent behaviour or format that prompting can't reliably achieve; you want to cut cost or latency by using a smaller fine-tuned model or shorter prompts at high volume; or you have a specialised task with lots of quality examples where a tuned model clearly beats prompting on your eval.

Fine-tuning should be justified by measured gains.

Training a smaller 'student' model on outputs from a larger 'teacher' model to get similar task performance at lower cost.

A common reason companies fine-tune small models.

Training a model using pairs of preferred vs rejected responses so it learns which outputs people prefer.

Shapes quality and style beyond imitation.

Add frequently changing factual knowledge

Facts baked into weights go stale and can be hallucinated; use retrieval instead.

Trains small low-rank adapter matrices instead of all weights

Parameter-efficient fine-tuning: cheaper to train and store.

A clear task definition and an eval set with a prompting baseline to beat; enough high-quality, diverse, representative training examples (cleaned and deduplicated, with no eval leakage); a budget and plan for training and hosting; and a process to retrain as data or requirements change.

Data quality matters more than quantity.

Explain that fine-tuning is unreliable for injecting facts: it can't cite sources, goes stale as the wiki changes, may hallucinate mixed-up facts, and is expensive to update. Recommend RAG over the wiki (with citations and access controls) and, if needed, fine-tuning only for style or format. Propose a quick prototype and eval to compare.

A classic interview scenario testing judgment.

Catastrophic forgetting

Mitigate with mixed training data, smaller learning rates, or adapters.