AI Pro - Digital Skill Cup 2026

Welcome to the Digital Skill Cup 2026 for AI Pro

Once you register with your name and email, the quiz starts immediately.
You have 5 minutes to complete the quiz.
No pauses. No switching between windows, no copy-pasting.
Scoring is based on the number of correct answers and the lower time to complete the quiz, in case of a tie.
Use a valid email address and your real name for registration.
You have 1 attempt.
Good luck :)

Time Remaining:

Digital Skill Cup 2026 - AI Pro All

1 / 20

Question 1

When switching to a different embedding model, what must you plan for?

Re‑indexing/re‑embedding the corpus because vector spaces are model‑specific.

Only updating metadata, since numeric vectors remain compatible.

No changes; all embedding models share the same vector space.

Regenerating citations but not the index.

Question 2

When deploying LLMs in production, which techniques are most effective at mitigating hallucinations without retraining the base model?

Increasing the temperature parameter to 1.5 to encourage more random outputs

Implementing RAG to ground responses in external data

Reducing the context window size to limit input complexity

Using system prompts to explicitly instruct the model to admit ignorance rather than hallucinating

Question 3

Which metric best measures agent workload without harming customer expirience for customer support chatbot?

Total conversations started

Exact match intent accuracy on a test set

Deflection rate combined with post chat CSAT for deflected sessions

Average handle time for human agents only

Question 4

Which prerequisite is required to enable Copilot for Microsoft 365 in an organization?

A personal Microsoft account

Microsoft 365 E3 or E5 licensing (or equivalent)

A local Active Directory only

A Windows Home license

Question 5

In Azure Machine Learning, what is the specific purpose of a Compute Instance compared to a Compute Cluster?

A Compute Instance is used exclusively for distributed training of deep learning models, whereas a Compute Cluster is for debugging

A Compute Instance is a fully configured development environment in the cloud used for authoring and testing code, whereas a Compute Cluster is used for scalable batch processing and training jobs

A Compute Instance is serverless and auto-scales based on HTTP traffic, whereas a Compute Cluster is a fixed set of virtual machines

A Compute Instance cannot access storage accounts, whereas a Compute Cluster has native access to Azure Blob Storage

Question 6

What is a key benefit of transfer learning?

Removes bias

Requires more data

Eliminates testing

Reduces training time

Question 7

In Azure AI Search: What is the advantage of "Vector Search" compared to classic full-text search?

It requires no indexing

It finds results based on semantic similarity rather than just keywords

It is always free

It only works with image files

Question 8

What type of data does Microsoft Copilot use to generate responses?

Public internet data only

Organizational data and Microsoft Graph (with permissions)

Encrypted passwords

Raw database files

Question 9

In Transformer architectures, what is the primary benefit of Multi-Head Attention compared to single-head attention?

To sequentially process data from left to right to maintain temporal order

To allow the model to attend to different parts of the input sequence simultaneously, capturing long-range dependencies

To reduce the dimensionality of the input embedding layer via principal component analysis

To apply non-linear activation functions recursively to improve gradient flow

Question 10

When using feature scaling or normalization in a machine learning pipeline (e.g., with distance-based algorithms or neural networks), what is the correct procedure to avoid data leakage?

Apply scaling to the entire dataset before splitting into training and test sets

Split the data first, fit the scaler on the training set only, then apply it to validation/test sets

Apply scaling independently to each dataset split using their own statistics

Feature scaling is only required for algorithms that use gradient descent

Question 11

Which transformer feature enables handling long-range dependencies?

Convolution

Pooling

Dropout

Attention mechanism

Question 12

What experiment best yields reliable conclusions when comparing two LLM prompts A and B for sales email generator?

Ask colleagues to pick their favorite outputs

Generate 10 samples per prompt and pick the one that reads best for you

Use offline BLEU score against a reference email

Run an A/B test with randomized assignment

Question 13

On AWS, which architecture us mos appropriate for scalable low latency RAG application over internal knowledge bases?

Amazon Bedrock for foundation models + Amazon OpenSearch with vector search + Amazon API Gateway/Lambda

AWS Glue + AWS Batch

Amazon SageMaker JumpStart + Amazon S3 only

Amazon EMR + Amazon Kinesis Data Firehose

Question 14

How does the vanishing gradient problem most directly impact the training of very deep neural networks?

It causes gradients to grow exponentially in deeper layers

It increases the memory footprint during deployment

It leads to unstable loss values during inference

It prevents effective learning in layers close to the input

Question 15

Which technique is most effective for reducing overfitting in deep neural networks?

Increasing the learning rate

Adding more hidden layers

Applying dropout regularization

Using a smaller dataset

Question 16

Which AWS AI service automates computer vision tasks like object and scene detection without model training?

Amazon Comprehend

Amazon SageMaker Ground Truth

Amazon Rekognition

Amazon Translate

Question 17

What is the main advantage of using transfer learning in AI models?

Eliminates the need for labeled data

Reduces training time and required data

Guarantees higher accuracy

Removes model bias completely

Question 18

Which of the following is NOT a common activation function in neural networks?

Sigmoid

ReLU

Linear

Euclidean

Question 19

Which evaluation approach best measures factual faithfulness of AI summaries?

Counting the number of citations regardless of quality.

Human‑in‑the‑loop review with source‑grounded rubrics plus targeted automated checks for claim‑citation alignment.

Perplexity of the generated text.

BLEU scores against a single reference summary.

Question 20

What does the term "RAG" (Retrieval-Augmented Generation) describe?

A method for image compression

Extending a model by retrieving external, up-to-date data during a query

Automatically deleting outdated training data

A special GPU cluster from NVIDIA