Concepts

Question 1

Concepts

What’s an n-gram?

Answer 1

1-gram is 1 word. 5-gram is 5 words, etc.

Question 2

Concepts

What is Tokenization?

Answer 2

Convert raw text into sequence of tokens

Question 3

Concepts

What happens to punctuation in Tokenization?

Answer 3

Kept. Likely each is a token

Question 4

Concepts

Is a word just a token?

Answer 4

Not necessarily – some words can split into mulitple Tokens

Question 5

Concepts

Example of a single word split into multiple tokens?

Answer 5

Richard might be “Rich” + “ard” because Rich and Richard are semantically close

Question 6

Concepts

What is a Context Window?

Answer 6

Max tokens the model can consider at once.

Question 7

Concepts

How does Context Window affect chatbots and prompting?

Answer 7

Limits how much context and question can be input to the model.

Question 8

Concepts

How does Context Window affect image or video input?

Answer 8

Size of the image or video along with any text context or prompts

Question 9

Concepts

What’s the relation between tokens and vectors in a vector DB?

Answer 9

A single token (like “cat”) corresponds to a single vector of many float values

Question 10

Concepts

Why does a single token have a vector with many values?

Answer 10

Captures many types of semantic meaning, sentiment, syntactic role, …

Question 11

Concepts

Example of how a vector is useful for searching a RAG?

Answer 11

Find similar tokens by looking for similar vectors

Question 12

Concepts

What’s the techncial name for this similarity search between vectors?

Answer 12

k-nearest neighbor

Question 13

Hyperparameters

What are hyperparameters?

Answer 13

Settings that define the model structure, algorithm and process

Question 14

Hyperparameters

When do you set hyperparameters?

Answer 14

Have to be set before you tune – hyperparameters define how it learns

Question 15

Hyperparameters

Example hyperparameters?

Answer 15

Learning rate, batch size, number of epochs

Question 16

Hyperparameters

What do you get by tuning hyperparameters?

Answer 16

Reduce overfitting, improve accuracy

Question 17

Hyperparameters

What is Learning Rate?

Answer 17

How large or small the steps are when updating weights during training

Question 18

Hyperparameters

What happens if you set a small learning rate?

Answer 18

Takes a long time to converge, get accurate weights

Question 19

Hyperparameters

What happens if you set a high learning rate?

Answer 19

Quicker to converge, but could over-shoot the right weights and not be accurate enough

Question 20

Hyperparameters

What is Batch Size?

Answer 20

Number of examples used to update a model weight in one iteration

Question 21

Hyperparameters

What happens with small batch sizes?

Answer 21

More stable learning, but takes more time to compute

Question 22

Hyperparameters

What happens with large batch sizes?

Answer 22

Faster, but unstable – too much info can swing weights violently.

Question 23

Hyperparameters

What is Number of Epochs?

Answer 23

How many times the model will iterate over the entire training set

Question 24

Hyperparameters

What happens with too few Epochs?

Answer 24

Underfitting

Question 25

# *Hyperparameters* What happens with too many Epochs?

Answer 25

Overfitting

Question 26

# *Hyperparameters* What happens with training data set is too small?

Answer 26

Overfitting: no outlyers, no noise, so everything fits (too) well

Question 27

# *Hyperparameters* How do you prevent overfitting?

Answer 27

Large training data set size, fewer epochs