Generative AI Leader Practice Question: A data scientist is fine-tuning a foundation…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✓

  Low-Rank Adaptation (LoRA)

  Why this is correct

  Low-Rank Adaptation freezes the pretrained weights and injects trainable rank-decomposition matrices into each layer, so only a small fraction of parameters update. With just 5,000 examples, this slashes computational cost and memory while preserving the original weights, directly preventing catastrophic forgetting.

• ✗

  Reinforcement Learning from Human Feedback (RLHF)

  Why it's wrong here

  RLHF optimises model outputs against a reward model trained on human preference rankings; it does not adapt the model to a labelled summarisation corpus. It suits aligning an already capable model with human preferences. With 5,000 examples, the stem calls for parameter-efficient supervised fine-tuning such as LoRA.

• ✗

  Full supervised fine-tuning of all model parameters

  Why it's wrong here

  Updating every weight with 5,000 examples overwrites the pretrained representations that generalise, producing catastrophic forgetting, and requires GPU memory and compute for full-model gradients and optimiser states. Full fine-tuning suits large corpora with ample hardware. The stem's constraints point to parameter-efficient methods such as LoRA.

• ✗

  In-context learning with few-shot examples

  Why it's wrong here

  In-context learning injects examples into the prompt at inference; it updates no weights, so it cannot adapt the model to legal summarisation or prevent forgetting. It suits rapid prototyping when no training budget exists. The stem requires parameter-efficient fine-tuning, such as LoRA, which trains small adapter matrices.

Go deeper

Related to this question