LLM — da5004_2025T2_Q2_NA

Source: da5004_2025T2_Q2_NA.pdf · Academic questions: 17 · Separately counted children: 0
Provenance: Large_Language_Models/da5004_2025T2_Q2_NA.pdf · SHA-256 86c404a53c5610cb48b7c7d701bb4794a5f76fe4fb77a9aa07df66976036a9a5

Scope: Large Language Models (LLM) section (Section 13, Section Id: 64065394491) spanning page 1 at y=78.10 to page 9 at y=842.00. Preceding Section 12 tail content at top of page 1 (y=0 to y=78.10, Possible Answers : 2) is excluded as out of subject scope.

Questions

Question 269

Question ID: 6406531326571

Type: MCQ · Marks: 2
PDF pages: 2

What is a primary motivation for using subword tokenization (like BPE or WordPiece) instead of
word-level tokenization in transformers?

Options

It increases vocabulary size exponentially

Source option ID: 6406534441547

It completely eliminates the need for a tokenizer

Source option ID: 6406534441548

It helps handle out-of-vocabulary words more effectively

Source option ID: 6406534441549

It prevents the model from needing positional encoding

Source option ID: 6406534441550

Correct answer: It helps handle out-of-vocabulary words more effectively

Source evidence: [{"page": 2, "bbox": [23.62, 191.38, 131.02, 207.7], "type": "marked-options", "detail": "Green checkmark option 6406534441549"}]

Question 270

Question ID: 6406531326572

Type: MCQ · Marks: 2
PDF pages: 2

What is a key implication of scaling laws in large language models?

Options

Training longer always leads to better performance

Source option ID: 6406534441551

Larger models require less training data

Source option ID: 6406534441552

Model size and dataset size must scale together to maintain optimal
performance

Source option ID: 6406534441553

None of these

Source option ID: 6406534441554

Correct answer: Model size and dataset size must scale together to maintain optimal
performance

Source evidence: [{"page": 2, "bbox": [23.62, 395.08, 131.02, 411.4], "type": "marked-options", "detail": "Green checkmark option 6406534441553"}]

Question 271

Question ID: 6406531326573

Type: MCQ · Marks: 2
PDF pages: 2

Consider the Following Assertion and Reason pair :
Assertion : In Byte Pair Encoding (BPE), each merge operation adds exactly one new token to the
vocabulary.
Reason : BPE helps reduce the size of the initial vocabulary by merging the most frequent tokens.

Options

Both Assertion and Reason are true, and Reason is the correct explanation of
Assertion.

Source option ID: 6406534441555

Both Assertion and Reason are true, but Reason is not the correct explanation
of Assertion.

Source option ID: 6406534441556

Assertion is true, but Reason is false.

Source option ID: 6406534441557

Assertion is false, but Reason is true.

Source option ID: 6406534441558

Correct answer: Assertion is true, but Reason is false.

Source evidence: [{"page": 2, "bbox": [23.62, 696.7, 131.02, 713.02], "type": "marked-options", "detail": "Green checkmark option 6406534441557"}]

Question 272

Question ID: 6406531326575

Type: MCQ · Marks: 2
PDF pages: 2, 3

What is the primary advantage of T5’s text-to-text approach for transfer learning in NLP?

Options

It reduces the computational cost of training models significantly.

Source option ID: 6406534441563

It allows a single model and architecture to be applied to a wide variety of NLP
tasks without modification.

Source option ID: 6406534441564

It guarantees optimal performance on all tasks without fine-tuning.

Source option ID: 6406534441565

It eliminates the need for any pre-training data.

Source option ID: 6406534441566

Correct answer: It allows a single model and architecture to be applied to a wide variety of NLP
tasks without modification.

Source evidence: [{"page": 3, "bbox": [23.62, 76.96, 131.02, 93.28], "type": "marked-options", "detail": "Green checkmark option 6406534441564"}]

Question 273

Question ID: 6406531326576

Type: MCQ · Marks: 2
PDF pages: 3

What is the format of output the T5 model is expected to generate during its span-corruption
pretraining?

Options

A sentence with masked tokens filled in

Option A: A sentence with masked tokens filled in

Figure 1. Option A: A sentence with masked tokens filled in

Source option ID: 6406534441567

A concatenation of masked spans in order, each preceded by the corresponding sentinel token

Option B: A concatenation of masked spans in order, each preceded by the corresponding sentinel token

Figure 2. Option B: A concatenation of masked spans in order, each preceded by the corresponding sentinel token

Source option ID: 6406534441568

A binary classification of whether the span was masked correctly

Option C: A binary classification of whether the span was masked correctly

Figure 3. Option C: A binary classification of whether the span was masked correctly

Source option ID: 6406534441569

A sequence of original and corrupted spans separated by <mask>

Option D: A sequence of original and corrupted spans separated by <mask>

Figure 4. Option D: A sequence of original and corrupted spans separated by <mask>

Source option ID: 6406534441570

Correct answer: A concatenation of masked spans in order, each preceded by the corresponding sentinel token

Source evidence: [{"page": 3, "bbox": [23.62, 362.48, 131.02, 378.8], "type": "marked-options", "detail": "Green checkmark option 6406534441568"}]

Question 274

Question ID: 6406531326574

Type: MCQ · Marks: 3
PDF pages: 3, 4

Given the input string:
peanutbutter

And the following vocabulary of subword tokens with their corresponding log-probabilities:

Subword	Log-Probability (base \(e\))
pea	-1.0
nut	-0.8
but	-1.2
ter	-0.7
butter	-1.0
peanut	-0.5
t	-2.0
er	-1.5

Using the Viterbi algorithm (as used in the SentencePiece tokenizer), determine the most probable tokenization of the input string. The probability of a tokenized sequence is the sum of the log-probabilities of the selected subwords. Only subwords from the vocabulary may be used.

Viterbi algorithm tokenization problem with input string peanutbutter and vocabulary table

Figure 5. Viterbi algorithm tokenization problem with input string peanutbutter and vocabulary table

Options

[peanut, but, ter]

Option A: [peanut, but, ter]

Figure 6. Option A: [peanut, but, ter]

Source option ID: 6406534441559

[pea, nut, butter]

Option B: [pea, nut, butter]

Figure 7. Option B: [pea, nut, butter]

Source option ID: 6406534441560

[peanut, butter]

Option C: [peanut, butter]

Figure 8. Option C: [peanut, butter]

Source option ID: 6406534441561

[pea, nut, but, ter]

Option D: [pea, nut, but, ter]

Figure 9. Option D: [pea, nut, but, ter]

Source option ID: 6406534441562

Correct answer: [peanut, butter]

Source evidence: [{"page": 4, "bbox": [23.62, 471.4, 131.02, 487.72], "type": "marked-options", "detail": "Green checkmark option 6406534441561"}]

Question 275

Question ID: 6406531326577

Type: MCQ · Marks: 3
PDF pages: 4, 5

Suppose you are given 4x more compute budget to scale a Transformer-based language model.
According to scaling laws, what is the best way to utilize this increased compute for optimal
model performance?

Options

Increase the model size (number of parameters) by 4x and keep the dataset
size and training steps the same.

Source option ID: 6406534441571

Train the same model on 4x more data for 4x more steps without changing
the model size.

Source option ID: 6406534441572

Increase the model size by approximately 2x and increase the amount of
training data by approximately 2x.

Source option ID: 6406534441573

Double the model size and halve the training data size to compensate.

Source option ID: 6406534441574

Correct answer: Increase the model size by approximately 2x and increase the amount of
training data by approximately 2x.

Source evidence: [{"page": 4, "bbox": [23.62, 781.21, 131.02, 797.53], "type": "marked-options", "detail": "Green checkmark option 6406534441573"}]

Question 276

Question ID: 6406531326578

Type: MSQ · Marks: 3
PDF pages: 5

In a data preprocessing stage designed to prepare clean, diverse, and safe data, which of the
following examples is likely to be filtered out?

Options

“All people from that xyz area are thieves. They should be jailed immediately.”

Source option ID: 6406534441575

Kavya’s PAN number is BDZPK1234K and her registered address is 17/4, Green
Park, Delhi.

Source option ID: 6406534441576

[“Reinforcement learning allows agents to learn optimal strategies through
interaction with an environment.”, “Reinforcement learning allows agents to learn optimal
strategies through interaction with an environment.”]

Source option ID: 6406534441577

The Union Budget 2024 proposes increased spending on green energy
infrastructure.

Source option ID: 6406534441578

According to WHO, global life expectancy has increased by 6 years since 2000.

Source option ID: 6406534441579

None of these.

Source option ID: 6406534441580

Correct answer: “All people from that xyz area are thieves. They should be jailed immediately.”; Kavya’s PAN number is BDZPK1234K and her registered address is 17/4, Green
Park, Delhi.; [“Reinforcement learning allows agents to learn optimal strategies through
interaction with an environment.”, “Reinforcement learning allows agents to learn optimal
strategies through interaction with an environment.”]

Source evidence: [{"page": 5, "bbox": [23.62, 233.02, 131.02, 249.34], "type": "marked-options", "detail": "Green checkmark option 6406534441575"}, {"page": 5, "bbox": [23.62, 253.84, 131.02, 270.16], "type": "marked-options", "detail": "Green checkmark option 6406534441576"}, {"page": 5, "bbox": [23.62, 290.98, 131.02, 307.3], "type": "marked-options", "detail": "Green checkmark option 6406534441577"}]

Question 277

Question ID: 6406531326579

Type: MSQ · Marks: 3
PDF pages: 5

What are the key challenges encountered when building a vocabulary for NLP models?

Options

A larger vocabulary leads to a bigger embedding matrix and increased
softmax computation

Source option ID: 6406534441581

Difficulty in handling out-of-vocabulary (OOV) words

Source option ID: 6406534441582

The size of each embedding vector increases with the number of out-of-
vocabulary (OOV) words

Source option ID: 6406534441583

More vocabulary entries always guarantee better model generalization

Source option ID: 6406534441584

Correct answer: A larger vocabulary leads to a bigger embedding matrix and increased
softmax computation; Difficulty in handling out-of-vocabulary (OOV) words

Source evidence: [{"page": 5, "bbox": [23.62, 543.64, 131.02, 559.96], "type": "marked-options", "detail": "Green checkmark option 6406534441581"}, {"page": 5, "bbox": [23.62, 580.78, 131.02, 597.1], "type": "marked-options", "detail": "Green checkmark option 6406534441582"}]

Question 278

Question ID: 6406531326580

Type: MSQ · Marks: 3
PDF pages: 5, 6

Select all the correct input-output representations used for different pretraining objectives based
on the base sentence:

“The quick brown fox jumps over the lazy dog”

Options

Span Corruption (T5’s Objective)
Input: "The fox jumps over the dog"
Output: " quick brown lazy "

Option A: Span Corruption (T5 Objective) input-output representation

Figure 10. Option A: Span Corruption (T5 Objective) input-output representation

Source option ID: 6406534441585

Causal Language Modeling (CLM)
Input: "The quick brown fox jumps over the lazy"
Output: "The quick brown fox jumps over the lazy dog"

Option B: Causal Language Modeling (CLM) input-output representation

Figure 11. Option B: Causal Language Modeling (CLM) input-output representation

Source option ID: 6406534441586

Masked Language Modeling (MLM)
Input: "The [MASK] fox jumps over the lazy dog"
Output: "quick brown"

Option C: Masked Language Modeling (MLM) input-output representation

Figure 12. Option C: Masked Language Modeling (MLM) input-output representation

Source option ID: 6406534441587

Prefix Language Modeling
Input: "finish the sentence: The quick brown fox jumps"
Output: "over the lazy dog"

Option D: Prefix Language Modeling input-output representation

Figure 13. Option D: Prefix Language Modeling input-output representation

Source option ID: 6406534441588

Correct answer: Span Corruption (T5’s Objective)
Input: "The fox jumps over the dog"
Output: " quick brown lazy "; Prefix Language Modeling
Input: "finish the sentence: The quick brown fox jumps"
Output: "over the lazy dog"

Source evidence: [{"page": 6, "bbox": [23.62, 58.68, 131.02, 75.0], "type": "marked-options", "detail": "Green checkmark option 6406534441585"}, {"page": 6, "bbox": [23.62, 302.65, 131.02, 318.97], "type": "marked-options", "detail": "Green checkmark option 6406534441588"}]

Question 279

Question ID: 6406531326581

Type: SA · Marks: 2
PDF pages: 6, 7

In multi-task pretraining, tasks are sampled according to the Proportional Mixing Strategy:

\[r_m = \frac{\min(e_m, K)}{\sum_{n=1}^N \min(e_n, K)}\]

where:

\(e_m\) is the original number of examples in task \(m\),
Say we have \(N\) tasks and original number of samples in each task is \(e_n\) where \(n \in \{1, 2, \dots, N\}\)
\(K\) is the cap applied to all datasets.

Suppose you have two tasks:

Task A: \(e_1 = 10,000\)
Task B: \(e_2 = 1,000,000\)

What is the largest value of \(K\) such that neither task has a sampling probability more than 50% (i.e. \(\le 0.5\))?

Multi-task pretraining Proportional Mixing Strategy mathematical formulation and problem

Figure 14. Multi-task pretraining Proportional Mixing Strategy mathematical formulation and problem

Correct answer: 10000

Source evidence: [{"page": 7, "bbox": [22.9, 113.4, 132.2, 151.0], "type": "numeric-answer", "detail": "Possible Answers : 10000"}]

Question 280

Question ID: 6406531326582

Type: SA · Marks: 3
PDF pages: 7

Consider the scaling law for pretraining test loss:

\[L(N, D) = \left(\frac{N_c}{N}\right)^{\alpha_N} + \left(\frac{D_c}{D}\right)^{\alpha_D}\]

where \(A = \left(\frac{N_c}{N}\right)^{\alpha_N}\), \(B = \left(\frac{D_c}{D}\right)^{\alpha_D}\)

\(N\): the number of non-embedding parameters
\(D\): the size of the dataset (in tokens)
\(N_c = 8.8 \times 10^8\)
\(D_c = 5.4 \times 10^9\)
\(\alpha_N = 0.076\)
\(\alpha_D = 0.095\)

Assume \(A = 2B\). What is the approximate ratio of the new loss after doubling \(N\) and \(D\) to the original loss i.e., \(\frac{L(2N, 2D)}{L(N, D)}\)? (Answer correct upto 3 digits after the decimal)

Scaling law for pretraining test loss mathematical formulation and parameter values

Figure 15. Scaling law for pretraining test loss mathematical formulation and parameter values

Correct answer: 0.914 to 0.975

Source evidence: [{"page": 7, "bbox": [22.9, 750.9, 132.2, 788.6], "type": "numeric-answer", "detail": "Possible Answers : 0.914 to 0.975"}]

Shared context 281 to 285

Question ID: 6406531326583

Portal container ID: 6406531326583

Type: Comprehension
PDF pages: 8

Consider the following dictionary for the subword tokenization:

vocabulary = {
    "gentler": 4,
    "brighter": 5,
    "crisper": 3
}

Append </w> to each word and consider it for all the calculations.

Based on the above data, answer the given subquestions.

Vocabulary dictionary for subword tokenization with gentler, brighter, crisper and frequency counts

Figure 16. Vocabulary dictionary for subword tokenization with gentler, brighter, crisper and frequency counts

Question 281

Question ID: 6406531326584

Type: SA · Marks: 2
PDF pages: 8

Shared context: ctx-281-285

Using Byte Pair Encoding (BPE), how many distinct tokens are there in the initial vocabulary before
any merges are performed?

Correct answer: 13

Source evidence: [{"page": 8, "bbox": [22.9, 562.5, 132.2, 600.2], "type": "numeric-answer", "detail": "Possible Answers : 13"}]

Question 282

Question ID: 6406531326585

Type: MCQ · Marks: 2
PDF pages: 8

Shared context: ctx-281-285

Which token pair has the lowest frequency before any merges in Byte Pair Encoding?

Options

(‘b’, ‘r’)

Source option ID: 6406534441592

(‘t’, ‘l’)

Source option ID: 6406534441593

(‘p’, ‘e’)

Source option ID: 6406534441594

(‘e’, ‘r’)

Source option ID: 6406534441595

None of these

Source option ID: 6406534441596

Correct answer: (‘p’, ‘e’)

Source evidence: [{"page": 8, "bbox": [23.62, 761.46, 131.02, 777.78], "type": "marked-options", "detail": "Green checkmark option 6406534441594"}]

Question 283

Question ID: 6406531326586

Type: SA · Marks: 2
PDF pages: 9

Shared context: ctx-281-285

What is the frequency of the most frequent byte-pair before the first merge in Byte Pair Encoding?

Correct answer: 12

Source evidence: [{"page": 9, "bbox": [22.9, 210.0, 132.2, 247.6], "type": "numeric-answer", "detail": "Possible Answers : 12"}]

Question 284

Question ID: 6406531326587

Type: SA · Marks: 2
PDF pages: 9

Shared context: ctx-281-285

If after three merges using Byte Pair Encoding (BPE), the vocabulary is as shown below, how many subword tokens will be used to tokenize the word “writer” ?

['</w>', 'b', 'c', 'e', 'g', 'h', 'i', 'l', 'n', 'p', 'r', 's', 't', 'er', 'er</w>', 'ri']

BPE vocabulary after three merges and tokenization question for word writer

Figure 17. BPE vocabulary after three merges and tokenization question for word writer

Correct answer: 4

Source evidence: [{"page": 9, "bbox": [22.9, 545.8, 132.2, 583.4], "type": "numeric-answer", "detail": "Possible Answers : 4"}]

Question 285

Question ID: 6406531326588

Type: MCQ · Marks: 2
PDF pages: 9

Shared context: ctx-281-285

Which of the following pairs will be merged in the first iteration based on the following score used in the WordPiece tokenizer?

\[\text{score} = \frac{\text{count}(\alpha, \beta)}{\text{count}(\alpha) \times \text{count}(\beta)}\]

WordPiece tokenizer scoring formula and pair merging question

Figure 18. WordPiece tokenizer scoring formula and pair merging question

Correct answer: Answer not verifiable from the source PDF

Source evidence: [{"page": 9, "bbox": [22.87, 663.72, 278.62, 807.3], "type": "absent", "detail": "Question 285 ends after Options : on final page 9 of the source PDF with no option items or official answer key provided"}]

Answer key

Source question	Source identity	Correct answer	PDF pages
269	6406531326571	It helps handle out-of-vocabulary words more effectively	2
270	6406531326572	Model size and dataset size must scale together to maintain optimal performance	2
271	6406531326573	Assertion is true, but Reason is false.	2
272	6406531326575	It allows a single model and architecture to be applied to a wide variety of NLP tasks without modification.	2, 3
273	6406531326576	A concatenation of masked spans in order, each preceded by the corresponding sentinel token	3
274	6406531326574	[peanut, butter]	3, 4
275	6406531326577	Increase the model size by approximately 2x and increase the amount of training data by approximately 2x.	4, 5
276	6406531326578	“All people from that xyz area are thieves. They should be jailed immediately.”; Kavya’s PAN number is BDZPK1234K and her registered address is 17/4, Green Park, Delhi.; [“Reinforcement learning allows agents to learn optimal strategies through interaction with an environment.”, “Reinforcement learning allows agents to learn optimal strategies through interaction with an environment.”]	5
277	6406531326579	A larger vocabulary leads to a bigger embedding matrix and increased softmax computation; Difficulty in handling out-of-vocabulary (OOV) words	5
278	6406531326580	Span Corruption (T5’s Objective) Input: "The fox jumps over the dog" Output: " quick brown lazy "; Prefix Language Modeling Input: "finish the sentence: The quick brown fox jumps" Output: "over the lazy dog"	5, 6
279	6406531326581	10000	6, 7
280	6406531326582	0.914 to 0.975	7
281	6406531326584	13	8
282	6406531326585	(‘p’, ‘e’)	8
283	6406531326586	12	9
284	6406531326587	4	9
285	6406531326588	Answer not verifiable from the source PDF	9

Extraction notes

Status: verified-source-limitation
Count reconciliation: Section 13 declares 14 portal questions and 40 section marks. The 14 portal containers comprise: 1 administrative confirmation question (Q268, 0 marks, excluded), 12 standalone questions (Q269: 2 marks, Q270: 2 marks, Q271: 2 marks, Q272: 2 marks, Q273: 2 marks, Q274: 3 marks, Q275: 3 marks, Q276: 3 marks, Q277: 3 marks, Q278: 3 marks, Q279: 2 marks, Q280: 3 marks; total standalone = 30 marks), and 1 comprehension portal container (Q281-285: 5 subquestions of 2 marks each = 10 marks). Total academic questions = 12 standalone + 5 comprehension subquestions = 17 academic questions. Total marks = 30 + 10 = 40 marks, exactly reconciling the declared section marks and 14 portal containers (1 confirmation + 12 standalone + 1 comprehension container = 14 portal questions).

Section 13 LLM examination starts on page 1 after preceding section tail and continues through page 9.
Question 285 (MCQ, 2 marks, Question Id: 6406531326588) on Page 9 has stem header and stem image (WordPiece tokenizer scoring formula) followed by 'Options :', but no option items or official answer key are present in the source PDF before the document ends at page 9. In accordance with extraction contract instructions, the answer is recorded with kind 'absent' and text 'Answer not verifiable from the source PDF'.

Excluded source entries:

{"classification": "out-of-scope-tail", "page": 1, "bbox": [22.87, 16.8, 169.2, 54.4], "reason": "Tail of preceding Section 12 answer evaluation ('Possible Answers : 2') appearing above Section 13 LLM header", "identity": "section-12-tail"}
{"classification": "subject-confirmation", "page": 1, "bbox": [22.87, 396.6, 580.0, 765.0], "reason": "Administrative subject confirmation asking candidate to confirm subject and cross check hall ticket", "identity": "268"}