Fine-Tuning & Model Adaptation

Practise and Check

Deliverable

RAG or fine-tune: for one narrow task, compare the M10 RAG baseline with a small fine-tuned model on the same questions and write down which wins and why.

Assessment

Answer the five Interview questions without notes, then run the lab.

Interview probe

Larkspur fine-tuned a small model on 2,000 support transcripts. In testing it answers the refund question correctly, then keeps going: it repeats the policy sentence until the response hits its token limit. What do you check?

Common mistakes

• Fine-tuning facts that change, such as the refund window, instead of retrieving them
• Changing the prompt format at serving time but not in training
• Training on rows with no end token, so answers never stop

Related Modules

• T16.12Model adaptation and inference serving