Free AWS AIP-C01 Actual Exam Questions - Question 1 Discussion
Scenario: An email filtering system needs to fine-tune a pre-trained BERT model for spam detection using a labeled email dataset (binary classification). Goal: correctly load the pretrained weights and use them as the initialization point for fine-tuning without full retraining. Question- Which approach will correctly initialize the BERT model to achieve this requirement?. Options:
Option B looks right because the original BERT output layer is generally for masked language modeling, not classification. So, discarding that and putting a new classifier on top suits the spam detection task. Options A and C keep or add classifiers without removing the original head, which could mess up the fine-tuning. D tries to turn it into a multi-task classifier, which isn’t necessary here since the aim is just binary spam detection. Swapping out the final layer while keeping all pretrained transformer layers untouched seems like the cleanest way to fine-tune without retraining everythin
It’s B because you want to keep all pretrained layers intact but swap out the last layer for a classifier suited to spam detection. Adding a parallel classifier like in C seems overcomplicated here.
It’s A because you keep all pretrained layers untouched and just train the new classifier.
B/C? B is typical—replace final layer for task; C’s parallel layer seems unnecessary.
B/C? B seems standard—replace final layer. C sounds odd adding parallel classifier.
Maybe A works too because it keeps all pretrained layers intact and only trains the new classifier on top, avoiding full retraining while still leveraging learned features.
I’m thinking option A might not be ideal because it suggests just adding a classifier on top of the existing model output without replacing the final layer, which might not align well if the original was trained for a different task. C talks about adding a parallel classifier, which could confuse the model by having two heads at once—usually it’s better to replace the head fully. D seems off since multi-task learning isn’t the goal here. So between these, B seems more direct for fine-tuning by swapping out the final layer while keeping pretrained weights intact. Anyone else feel like A or C co
B, since you want to keep pretrained layers but swap the final classifier for spam detection.
A vs B, B sounds better since it replaces the final layer for the new task.
D imo, Kendra is built for semantic search and directly integrates with S3, which fits the scenario perfectly. The other options seem more about data extraction or custom setups, not optimized semantic search out of the box.
D seems like the best fit since Amazon Kendra is designed for semantic search and integrates directly with S3, making it simpler and more optimized for this use case.