Context window handling in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Understanding context window size impact

In natural language processing, what is the main effect of increasing the context window size when training a language model?

A The model can consider more words at once, improving understanding of long-range dependencies.

B The model trains faster because it processes fewer tokens per batch.

C The model reduces memory usage by limiting the number of tokens it stores.

D The model ignores earlier words and focuses only on the last few tokens.

💡 Hint

Think about how much text the model can see at once.

Increasing the context window size allows the model to see more words together, which helps it understand relationships between words that are far apart in the text.

❓ Predict Output

intermediate

Output of sliding window tokenization

What is the output of the following Python code that simulates a sliding context window over tokens?

NLP

tokens = ['I', 'love', 'machine', 'learning', 'and', 'AI']
window_size = 3
windows = [tokens[i:i+window_size] for i in range(len(tokens) - window_size + 1)]
print(windows)

A [['I', 'love', 'machine', 'learning'], ['love', 'machine', 'learning', 'and'], ['machine', 'learning', 'and', 'AI']]

B [['I', 'love', 'machine'], ['machine', 'learning', 'and'], ['love', 'machine', 'learning'], ['learning', 'and', 'AI']]

C [['I', 'love'], ['love', 'machine'], ['machine', 'learning'], ['learning', 'and'], ['and', 'AI']]

D [['I', 'love', 'machine'], ['love', 'machine', 'learning'], ['machine', 'learning', 'and'], ['learning', 'and', 'AI']]

💡 Hint

Look at how the list comprehension slices the tokens with a fixed window size.

The code creates overlapping windows of size 3 moving one token at a time, so each window contains 3 consecutive tokens.

❓ Model Choice

advanced

Choosing model type for long context handling

You want to build a model that can understand very long documents (thousands of words). Which model architecture is best suited to handle such long context windows efficiently?

A Transformer with sparse attention or memory-augmented mechanisms

B Recurrent Neural Network (RNN) with truncated backpropagation

C Standard Transformer with fixed 512-token context window

D Feedforward neural network without sequence modeling

💡 Hint

Consider models designed to reduce computation for long sequences.

Transformers with sparse attention or memory mechanisms can handle longer contexts by focusing attention on important parts, reducing computation compared to standard transformers.

❓ Hyperparameter

advanced

Effect of context window size on training time

How does increasing the context window size affect the training time of a Transformer-based language model?

A Training time decreases because the model sees more tokens at once.

B Training time increases roughly quadratically with context window size.

C Training time remains constant regardless of context window size.

D Training time increases linearly with context window size.

💡 Hint

Think about how attention computation scales with sequence length.

Self-attention in transformers computes relationships between all token pairs, so computation grows roughly with the square of the context window size.

🔧 Debug

expert

Identifying error in context window implementation

What error does the following code raise when trying to create context windows, and why?

tokens = ['a', 'b', 'c']
window_size = 4
windows = [tokens[i:i+window_size] for i in range(len(tokens) - window_size + 1)]
print(windows)

A IndexError because slicing goes beyond the list length.

B ValueError because window_size must be less than tokens length.

C Empty list because the range is negative, so no windows are created.

D TypeError because window_size is larger than tokens length.

💡 Hint

Check what happens when the range argument is negative.

The range becomes range(3 - 4 + 1) = range(0), which is empty, so windows is an empty list without error.