NEW
Font size
WorksheetsQuiz on Large Language Models and RAG
Total questions: 35
Worksheet time: 18mins
What is the defining characteristic of a Large Language Model (LLM)?
Only trained on code.
A deep neural network trained on a massive amount of text data.
Cannot generate new content.
Which of the following is not listed as a use case category for Generative AI in the materials?
Text and Code.
Purely administrative task logging.
Image and Data Augmentation.
The two key blocks in the original Transformer Architecture are:
Input and Output.
Encoder and Decoder.
Query and Key.
What is the key advantage of Parallel Processing in Transformers over the sequential processing used by RNNs?
It requires less training data.
It allows for much faster training and better contextual understanding.
It eliminates the need for any pre-training.
Which specific mechanism is used in the Decoder's self-attention layer to ensure it only attends to tokens that came before it?
Feed-Forward Masking.
Masked Self-Attention.
Positional Encoding.
What is the final step in the Full Fine-Tuning workflow after all weights have been trained and evaluated?
Chunking the data.
Save the updated model.
Rerunning the pre-training phase.
In the context of the Attention mechanism, what calculation determines how related each word is to the others?
Residual Connection.
Similarity Measure (e.g., dot product) between Query and Key.
Layer Normalization.
According to the materials, which model architecture is Encoder-only?
GPT.
BERT.
T5.
What is the primary benefit of Fine-Tuning a pre-trained LLM?
To decrease the number of parameters.
To adapt the model to a specific task, domain, or desired tone.
To convert the model to an Encoder-only structure.
What percentage of compute resources can PEFT techniques typically save compared to Full Fine-Tuning?
10.
90.
50.
Which of the following is not a specific PEFT technique mentioned in the materials?
LoRA.
Softmax Activation.
Adapter Tuning.
When is Full Fine-Tuning best suited?
When compute resources are severely limited.
For a new domain with a large proprietary dataset.
Only for multilingual models.
What term is used for the process of converting an LLM's prompt into a set of instructions using a declarative approach like YAML?
Vectorizing.
Prompt Declaration Language (PDL).
Transformer Activation.
In the Fine-Tuning workflow, after preparing the labeled dataset, what is the immediate next step?
Deploying the model.
Tokenize the data.
Performing a semantic search.
What is the role of Residual Connections within a Transformer block?
To add non-linearity.
To prevent vanishing gradients during training.
To compute the attention scores.
What is the purpose of Layer Normalization in a Transformer?
To select the final output token.
To stabilize training by normalizing the layer inputs.
To calculate the cosine similarity.
Which of the following is not one of the core components of the RAG architecture?
Retriever.
Data Tokenizer.
Knowledge Base.
The main idea behind RAG is a model that:
Generates images from text.
Looks up facts in an external knowledge base before generating a response.
Only answers questions based on its pre-training data.
Which step in the RAG process is where the user question is converted into a vector form?
Step 4: Final Generation.
Step 1: Query Understanding.
Step 3: Context Injection.
What is the benefit of using RAG over simply relying on the LLM's pre-training memory?
It makes the model smaller.
It enhances factual accuracy and provides more updated information.
It speeds up the model's training time by 90.
In RAG, what is the role of the Retriever component?
To generate the final output text.
To search the knowledge base/vector database for the most relevant documents.
To adjust the model's temperature parameter.
RAG is often described as a solution to the problem of LLMs having no access to:
The internet at all times.
Real-time or domain-specific data outside of their training data.
More than 100 parameters.
The RAG architecture is commonly referred to as the combination of:
Encoder and Decoder.
Retrieval and Generation loop.
Classification and Clustering.
Which step in RAG involves searching the Vector Database for the top-k relevant results?
Query Understanding.
Document Retrieval.
Final Answer Generation.
What is a major advantage of RAG in terms of model maintenance?
It allows the LLM to be retrained weekly.
It reduces the need for frequent LLM retraining when new data becomes available.
It replaces the need for embeddings entirely.
What is the key functional difference between BERT and GPT architectures, as noted in the materials?
BERT uses Multi-Head Attention, but GPT does not.
BERT is Encoder-only (Contextual Understanding), while GPT is Decoder-only (Generation).
BERT uses parallel processing, but GPT uses sequential.
According to the materials, the IBM Granite Model Series is optimized for:
Low-latency social media post generation.
Reliability, transparency, and domain adaptation for enterprise applications.
Exclusive use as a time-series forecasting model.
What must happen to a raw text document before it can be converted into an embedding?
It must be fully translated into one language.
It must be divided into smaller text 'chunks'.
It must be compressed into a zip file.
Why is Text Chunking necessary in the process of building a RAG pipeline?
To save on storage space in the Vector DB.
To fit content within the LLM's context window and maintain contextual continuity.
To increase the dimensionality of the resulting vector.
The core idea of Semantic Similarity (enabled by embeddings) is that:
All words are treated as unique, discrete keys.
Words or phrases with similar meaning will have vectors that are close in multidimensional space.
Only exact keyword matches will be retrieved from the database.
In the Vector DB context, a small angle between two vectors suggests:
Low similarity (high distance).
High similarity (close meaning).
The vectors are not indexed.
What critical information is typically stored alongside an embedding in a Vector Database to aid the RAG process?
The model's training objective.
Metadata (e.g., source, page number, topic).
The entire LLM parameter set.
What is the primary reason why models cannot process raw text directly?
Raw text is too long.
Models can only process numerical data.
Raw text is always proprietary.
What is one of the Core Operations in a Vector Database, besides searching and inserting?
Grammar Checking.
Indexing to optimize performance.
Generating a full fine-tuning dataset.
What is the term for the numerical form of meaning that captures context, tone, and semantics?
Tokenization.
Text Embedding.
Full Fine-Tuning.
