AIF-C01 Practice Question 8: Fine Tuning Output Behavior
AWS Certified AI Practitioner (AIF-C01) · Test 1 · question 8 of 10
A legal-tech company uses a foundation model to turn contract clauses into summaries that follow a strict house format and terminology. Getting acceptable results takes 25 worked examples in every prompt, which makes each request slow and expensive, and the output still drifts from the format. The required knowledge is in the clause itself and doesn't change. The company has 20,000 pairs of clauses and expert-approved summaries. Which approach best addresses both the cost and the quality problem?
Show answer and explanation
Answer:
The problem here is the model's behavior (format, style, terminology), not missing knowledge. The company also has a large labeled dataset of input-output pairs. That combination points to supervised fine-tuning. The model's weights learn the format directly, so the long run of in-prompt examples is no longer needed. Prompts get shorter, which lowers token cost and latency per request, and the output is usually more consistent than few-shot prompting can make it (option C). Fine-tuning has an upfront training cost, but at high volume, fewer tokens on every request pays it back.
- A is wrong because RAG is designed to supply missing or changing *facts*. Retrieving 25 examples still makes every prompt long, so the cost problem stays. - B is wrong because pre-training a foundation model from scratch costs vastly more than customizing an existing one, and it isn't needed to teach an output format. - D is wrong because more in-context examples *increase* tokens, cost, and latency on every request. That makes the company's main complaint worse.
Reference: Customize your model to improve its performance for your use case