uiz Space

January 2026 term · Large Language Models · BSDA5004

Large Language Models Quiz 2: 12 April 2026 (January 2026 term)

The IIT Madras BS Large Language Models (LLM) Quiz 2 paper sat on 12 Apr 2026, in the January 2026 term: 21 questions for 50 marks in 120 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.

Questions
21
Marks
50
Duration
120 min
MCQ
10
MSQ
6
Numerical
5

Updated

Official paper: Large Language Models 07 Apr 26 · No negative marking.

Question 1

+2 marksOne correct option
  1. A
  2. B
  3. C
  4. D
Show answer

Correct answer

  • C

Question 2

+2 marksOne correct option
  1. A

    To reduce the computational cost of the softmax layer during training.

  2. B
  3. C
  4. D

    To act as a regularizer similar to Dropout.

Show answer

Correct answer

  • C

Question 3

+2 marksOne correct option

Why is the standard GPT architecture (Decoder-only with causal masking) generally unsuitable for the Masked Language Modeling (MLM) objective as implemented in BERT?

  1. A

    GPT models are too small to learn bidirectional contexts.

  2. B

    The causal mask in GPT prevents the model from attending to future tokens, making it impossible to use right-side context to predict a masked token.

  3. C

    GPT does not have positional embeddings, which are required for MLM.

  4. D

    GPT uses ReLU activation, while BERT uses GELU, which is required for MLM.

Show answer

Correct answer

  • B

    The causal mask in GPT prevents the model from attending to future tokens, making it impossible to use right-side context to predict a masked token.

Question 4

+2 marksOne correct option

In the T5 (Text-to-Text Transfer Transformer) framework, every NLP task is cast as a text generation problem. If you use T5 for a Semantic Textual Similarity (STS-B) task, where the goal is to predict a similarity score (e.g., 3.8) between two sentences, how does the model output this score?

  1. A

    It outputs a single scalar value from a regression head on top of the encoder.

  2. B

    It generates the string "3.8" token-by-token using the decoder.

  3. C

    It outputs a class label corresponding to a bucketed score range (e.g., "High Similarity").

  4. D

    T5 cannot be used for regression tasks like STS-B.

Show answer

Correct answer

  • B

    It generates the string "3.8" token-by-token using the decoder.

Question 5

+2 marksOne correct option
  1. A

    The model trains almost exclusively on the largest dataset (highest resource task).

  2. B

    The sampling distribution approaches a uniform distribution, where all tasks (large and small) are sampled with nearly equal probability.

  3. C

    The model trains almost exclusively on the smallest dataset (lowest resource task).

  4. D

    The sampling distribution remains proportional to the original dataset sizes.

Show answer

Correct answer

  • B

    The sampling distribution approaches a uniform distribution, where all tasks (large and small) are sampled with nearly equal probability.

Question 6

+2 marksOne correct option

When constructing training datasets for large language models (LLMs), which of the following best describes the key factors that must be balanced to achieve strong and reliable performance?

  1. A

    Model depth, number of parameters, and learning rate

  2. B

    Scale, diversity, and quality of the training data

  3. C

    Vocabulary size, tokenization method, and batch size

  4. D

    Compute budget, optimizer choice, and hardware efficiency

Show answer

Correct answer

  • B

    Scale, diversity, and quality of the training data

Question 7

+3 marksOne or more correct options

Which of the following are components found within a standard Transformer Encoder layer?

Select all that apply.

  1. A

    Multi-Head Self-Attention mechanism

  2. B

    Position-wise Feed-Forward Networks

  3. C

    Cross-Attention mechanism (Encoder-Decoder attention)

  4. D

    Masked Multi-Head Self-Attention

Show answer

Correct answers

  • A

    Multi-Head Self-Attention mechanism

  • B

    Position-wise Feed-Forward Networks

Question 8

+3 marksOne or more correct options

Select all correct statements regarding the comparison between RNNs and Transformers.

Select all that apply.

  1. A
  2. B

    Transformers process tokens strictly sequentially during training in the same way as RNNs.

  3. C

    Transformers allow for significantly more parallelization during training compared to RNNs.

  4. D

    Attention mechanisms in Transformers utilize Query, Key, and Value vectors derived from input embeddings.

Show answer

Correct answers

  • A
  • C

    Transformers allow for significantly more parallelization during training compared to RNNs.

  • D

    Attention mechanisms in Transformers utilize Query, Key, and Value vectors derived from input embeddings.

Question 9

+3 marksOne or more correct options

Select all correct findings from the T5 paper regarding Unsupervised Pre-training Objectives.

Select all that apply.

  1. A

    The specific corruption rate (e.g., 10%, 15%, 25%) had a minimal effect on downstream performance.

  2. B

    Using a span length of around 3 tokens performed slightly better than masking single tokens.

  3. C

    The "Deshuffling" objective (reordering shuffled sentences) significantly outperformed the Span Corruption objective.

  4. D

    Replacing a corrupted span with a unique sentinel token worked better than simply dropping the tokens from the input.

Show answer

Correct answers

  • A

    The specific corruption rate (e.g., 10%, 15%, 25%) had a minimal effect on downstream performance.

  • B

    Using a span length of around 3 tokens performed slightly better than masking single tokens.

  • D

    Replacing a corrupted span with a unique sentinel token worked better than simply dropping the tokens from the input.

Question 10

+3 marksOne or more correct options

Consider the concept of Zero-Shot Transfer as popularized by GPT-2. Why might this be preferred over Supervised Fine-Tuning?

Select all that apply.

  1. A

    It allows the model to handle tasks for which no labeled training data is available.

  2. B

    It always achieves higher accuracy than a fine-tuned SOTA model.

  3. C

    It avoids the need to store a separate specialized model (checkpoint) for every downstream task.

  4. D

    It mimics the human ability to perform tasks based on instructions without needing thousands of examples.

Show answer

Correct answers

  • A

    It allows the model to handle tasks for which no labeled training data is available.

  • C

    It avoids the need to store a separate specialized model (checkpoint) for every downstream task.

  • D

    It mimics the human ability to perform tasks based on instructions without needing thousands of examples.

Question 11

+2 marksNumerical answer
Show answer

Correct answer: 1.335 (accepted within ±0.035)

Question 12

+2 marksOne correct option

Based on the above data, answer the given subquestions.

Is the given attention matrix valid for a causal language modelling task?

  1. A

    True

  2. B

    False

  3. C

    Insufficient information

Show answer

Correct answer

  • B

    False

Question 13

+2 marksNumerical answer

Based on the above data, answer the given subquestions.

At time step t = 5 (word: data), what is the attention weight assigned to the word science? (Provide exact answer)

Show answer

Correct answer: 0.2

Question 14

+2 marksOne correct option

Which of the following character pairs occurs with the highest frequency across the entire sentence before any BPE merges are performed?

  1. A
  2. B
  3. C
  4. D
Show answer

Correct answer

  • D

Question 15

+2 marksNumerical answer

Calculate the total vocabulary size immediately after the first merge operation is completed.
Note: The vocabulary size includes all individual base characters/symbols plus the newly created merge token.

Show answer

Correct answer: 20

Question 16

+3 marksNumerical answer
Show answer

Correct answer: 0.62 (accepted within ±0.03)

Question 17

+3 marksNumerical answer
Show answer

Correct answer: 648448

Question 18

+3 marksOne correct option

What is a major disadvantage of character-level tokenization?

  1. A

    Character-level tokenization cannot represent punctuation marks or special symbols

  2. B

    It has very large vocabulary size because each word is broken into multiple characters

  3. C

    It produces much longer input sequences which increases computational cost.

  4. D

    It fails to capture morphological patterns such as prefixes and suffixes, reducing the model's ability to understand word structure

Show answer

Correct answer

  • C

    It produces much longer input sequences which increases computational cost.

Question 19

+3 marksOne correct option

When using the SentencePiece, how are subword units selected?

  1. A

    Based on a probabilistic model that maximizes the likelihood of the training data.

  2. B

    By randomly selecting character n-grams until the vocabulary limit is reached.

  3. C

    By selecting only the top 50,000 most frequent words in the corpus.

  4. D

    By iteratively merging the most frequent pair of adjacent characters.

Show answer

Correct answer

  • A

    Based on a probabilistic model that maximizes the likelihood of the training data.

Question 20

+2 marksOne or more correct options

To obtain high-quality training text from raw web data for large language models, which of the following mechanisms are typically included in the pre-processing pipeline?

Select all that apply.

  1. A

    Tokenizing text into subword units using Byte Pair Encoding (BPE)

  2. B

    Deduplicating content at line, paragraph, and document levels

  3. C

    Detecting and filtering toxic content such as hate speech and profanity

  4. D

    Identifying the language of web pages

  5. E

    Assessing the quality of content to remove low-value or spam text

  6. F

    Detecting and removing Personally Identifiable Information (PII)

  7. G

    Fine-tuning the model using reinforcement learning from human feedback (RLHF)

Show answer

Correct answers

  • B

    Deduplicating content at line, paragraph, and document levels

  • C

    Detecting and filtering toxic content such as hate speech and profanity

  • D

    Identifying the language of web pages

  • E

    Assessing the quality of content to remove low-value or spam text

  • F

    Detecting and removing Personally Identifiable Information (PII)

Question 21

+2 marksOne or more correct options

Which of the following statements correctly explain the importance of deduplication during preprocessing of large-scale datasets used for training Deep Learning or Large Language Models?

Select all that apply.

  1. A

    Deduplication reduces the risk of overfitting by preventing repeated samples from dominating the gradient updates.

  2. B

    Deduplication guarantees that the trained model will achieve higher accuracy on all downstream tasks.

  3. C

    Deduplication helps avoid data leakage between training and evaluation sets, leading to more reliable performance metrics.

  4. D

    Deduplication eliminates the need for regularization techniques such as dropout and weight decay.

Show answer

Correct answers

  • A

    Deduplication reduces the risk of overfitting by preventing repeated samples from dominating the gradient updates.

  • C

    Deduplication helps avoid data leakage between training and evaluation sets, leading to more reliable performance metrics.