A bank is fine-tuning a large language model (LLM) on Amazon Bedrock to assist customers with questions about their loans. The bank wants to ensure that the model does not reveal any private…

Question

Options

• AUse Amazon Bedrock Guardrails.
• BRemove personally identifiable information (PII) from the customer data before fine-tuning the
• CIncrease the Top-K parameter of the LLM.
• DStore customer data in Amazon S3. Encrypt the data before fine-tuning the LLM.

• A
  9% (5)
• B
  72% (38)
• C
  4% (2)
• D
  15% (8)

Explanation

Explanation

Removing PII from training data before fine-tuning (Option B) is the most direct solution because it eliminates the risk at the source - if sensitive data never enters the model during training, the model fundamentally cannot learn to reproduce or reveal it, regardless of how users interact with it.

Why the distractors are wrong:

• Option A (Guardrails): While Bedrock Guardrails can help filter outputs, they are not foolproof and act as a reactive layer; if PII was baked into the model during training, it may still surface in unexpected ways.
• Option C (Top-K parameter): Top-K controls the diversity of token selection during text generation and has no relationship to data privacy or preventing PII disclosure.
• Option D (S3 encryption): Encrypting data at rest protects it from unauthorized storage access, but once that data is used to fine-tune the model, the model itself could still memorize and reproduce the PII.

Memory Tip: Think of it as "Garbage In, Garbage Out" - if you want a model to never output sensitive data, the cleanest strategy is to ensure that sensitive data was never fed in during training. Guardrails and encryption protect data around the model, but cleaning training data protects from within.

No community discussion yet for this question.