LLMs, Tokens and Embeddings Practice Test
Test how well you understand large language models: tokens, context windows, transformers, embeddings, fine-tuning and the settings developers use every day.
AI Skills
Before you start
- Questions
- 15
- Time
- 18 minutes
- Format
- Single section
- The timer starts when you press Start. When it reaches zero, your answers are submitted automatically.
- Use the numbered navigator to move between questions, and flag any you want to come back to.
- For typed answers, keep within the word limit. Answers over the limit are marked wrong, as in the real test.
- You can submit early. Your score, the correct answers and explanations appear straight away.
LLMs, Tokens and Embeddings Practice Test: all 15 questions with answers
Prefer to work through the test on paper? Answer the questions, then open “Show answer” to check.
Choose the correct answer, A, B, C or D.
Q1.What is a large language model mainly trained to do?
- A.Search the internet in real time
- B.Store a copy of every website
- C.Predict the next token given the tokens before it
- D.Run SQL queries
Show answer
Answer: C. Predict the next token given the tokens before it
LLMs are pre-trained on next-token prediction over very large text datasets.
Q2.What is a token?
- A.Always exactly one word
- B.A password for the API
- C.A single character only
- D.A chunk of text, such as a word or part of a word, that the model processes
Show answer
Answer: D. A chunk of text, such as a word or part of a word, that the model processes
Tokenisers split text into sub-word pieces. One English word is often about one to two tokens.
Q3.What is a model's context window?
- A.The maximum number of tokens the model can consider at once, including input and output
- B.The browser window where you chat
- C.The time limit for a request
- D.The number of users at once
Show answer
Answer: A. The maximum number of tokens the model can consider at once, including input and output
Anything beyond the context window cannot be seen by the model in that request.
Q4.Which mechanism is at the heart of the transformer architecture?
- A.Decision trees
- B.Self-attention
- C.K-means clustering
- D.Linear regression
Show answer
Answer: B. Self-attention
Self-attention lets each token weigh how relevant every other token is to it.
Q5.What is an embedding?
- A.A compressed image file
- B.A type of prompt
- C.A list of numbers (vector) that represents meaning, so similar texts end up close together
- D.A database table
Show answer
Answer: C. A list of numbers (vector) that represents meaning, so similar texts end up close together
Embeddings map text into vector space so meaning can be compared mathematically.
Q6.Which measure is commonly used to compare two embeddings?
- A.File size
- B.Word count
- C.Alphabetical order
- D.Cosine similarity
Show answer
Answer: D. Cosine similarity
Cosine similarity measures the angle between vectors: closer to 1 means more similar meaning.
Q7.What is fine-tuning?
- A.Further training a pre-trained model on task-specific data to adapt its behaviour
- B.Writing a longer prompt
- C.Lowering the temperature
- D.Deleting part of the model
Show answer
Answer: A. Further training a pre-trained model on task-specific data to adapt its behaviour
Fine-tuning adjusts the model's weights using a smaller, focused dataset.
Q8.How does pre-training differ from fine-tuning?
- A.They are the same thing
- B.Pre-training uses huge general datasets; fine-tuning uses smaller task-specific data
- C.Fine-tuning always uses more data
- D.Pre-training happens after deployment
Show answer
Answer: B. Pre-training uses huge general datasets; fine-tuning uses smaller task-specific data
Pre-training builds general language ability; fine-tuning specialises it.
Q9.Why might a model not know about a recent event?
- A.The model is broken
- B.The event was too short
- C.Its training data ends at a knowledge cutoff date
- D.Models never learn facts
Show answer
Answer: C. Its training data ends at a knowledge cutoff date
Models only know what was in their training data unless you supply newer information in the prompt or via tools.
Q10.What is RLHF used for?
- A.Compressing the model
- B.Translating code
- C.Aligning model outputs with human preferences using human feedback
- D.Speeding up tokenisation
Show answer
Answer: C. Aligning model outputs with human preferences using human feedback
Reinforcement learning from human feedback trains the model toward responses people rate as more helpful and safe.
Q11.Setting temperature to 0 usually makes output:
- A.More creative
- B.Longer
- C.Multilingual
- D.More deterministic and focused on the most likely tokens
Show answer
Answer: D. More deterministic and focused on the most likely tokens
Low temperature reduces randomness, which suits extraction, classification and other precise tasks.
Q12.Why does API cost usually depend on tokens?
- A.Providers charge per input and output token processed
- B.Tokens are physical items
- C.Cost depends only on the time of day
- D.Tokens are free
Show answer
Answer: A. Providers charge per input and output token processed
Most LLM APIs price usage by input and output tokens, so long prompts and long answers cost more.
Q13.What is a multimodal model?
- A.A model that only handles text
- B.A model that can process more than one type of input, such as text and images
- C.Several copies of the same model
- D.A model with several prompts
Show answer
Answer: B. A model that can process more than one type of input, such as text and images
Multimodal models accept and reason over different data types, for example a photo plus a question.
Q14.What does the max output tokens setting control?
- A.The size of the training data
- B.The number of users
- C.The maximum length of the generated response
- D.The temperature
Show answer
Answer: C. The maximum length of the generated response
If the limit is too low, the answer is cut off mid-way.
Q15.Why do chat apps stream responses token by token?
- A.It reduces the total cost to zero
- B.It is the only way models work
- C.It makes answers more accurate
- D.Users see text sooner, which improves perceived speed
Show answer
Answer: D. Users see text sooner, which improves perceived speed
Streaming lowers time-to-first-token for the user. It does not change the answer itself.