hardMultiple Choice
Generative AI Leader Practice Question: Fine-tuning a large language model for a…
A company is fine-tuning a large language model for a domain-specific legal document summarization task. They have limited labeled data but want to adapt the model efficiently without catastrophic forgetting. Which technique is most suitable?
⚠ Common exam trap
A common trap is the misconception that in-context learning (Option B) is sufficient for domain adaptation, but it does not modify model weights and thus cannot achieve the deep, consistent specialization required for tasks like legal document summarization.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Supervised fine-tuning using LoRA adapters
LoRA (Low-Rank Adaptation) is the most suitable technique because it enables parameter-efficient fine-tuning by injecting trainable low-rank matrices into the transformer layers, drastically reducing the number of updated parameters. This preserves the pre-trained knowledge and prevents catastrophic forgetting, even with limited labeled data, while efficiently adapting the model to domain-specific legal summarization.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Reinforcement Learning from Human Feedback (RLHF)
Why it's wrong here
RLHF optimises model behaviour against human preference rankings; it does not adapt a model to a narrow domain corpus and typically needs large volumes of preference annotations. It is tempting because RLHF is the standard technique for aligning general-purpose assistants, and would be correct when tuning tone, helpfulness or safety rather than legal summarisation.
- ✗
In-context learning with few-shot examples in prompts
Why it's wrong here
In-context learning leaves weights unchanged, so it cannot durably adapt the model to legal terminology and provides no protection against catastrophic forgetting because no parameter update occurs. It is tempting because few-shot prompting needs no labelled training data, and would be correct for quick prototyping or tasks already within the model's competence.
- ✓
Supervised fine-tuning using LoRA adapters
Why this is correct
LoRA freezes the base weights and trains small low-rank adapter matrices, so only a fraction of parameters update. This adapts efficiently under limited labelled data while preserving the pretrained knowledge that catastrophic forgetting would otherwise destroy.
- ✗
Full fine-tuning of all model parameters
Why it's wrong here
Updating every parameter with limited labelled legal data causes catastrophic forgetting of general language ability and overfits. It is tempting because full fine-tuning maximises task-specific accuracy, and would be correct when a large, high-quality domain corpus and substantial compute are available and general capability loss is acceptable.
Go deeper
Related to this question
About these practice questions
This Generative AI Leader question is part of Courseiva's 1,008-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This Generative AI Leader practice question is part of Courseiva's free Google Cloud certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the Generative AI Leader exam.