LLM — da5004_2026T2_Q2_NA

Source: da5004_2026T2_Q2_NA.pdf · Academic questions: 18 · Separately counted children: 0
Provenance: Large_Language_Models/da5004_2026T2_Q2_NA.pdf · SHA-256 f854547ec11886a82cb2287cf5e4c7d2b6f35b3aee3d17ad91f5cf8f48b47da0

Scope: Section 1: LLM (Degree Level : Large Language Models), physical pages 1 to 9. The in-scope section starts on page 1 with Section Id 640653143422 and concludes on page 9 with Question 19 under Comprehension 18 to 19.

Questions

Shared context 2-4

Question ID: 6406532197542

Type: Comprehension
PDF pages: 2

Answer the subquestions for the given transformer architecture for one ($N = 1$) Encoder Decoder Block

Transformer encoder-decoder architecture with labeled components for Comprehension questions 2 to 4

Figure 1. Transformer encoder-decoder architecture with labeled components for Comprehension questions 2 to 4

Question 2

Question ID: 6406532197543

Type: Short Answer Question · Marks: 2
PDF pages: 3

Shared context: ctx-2-4

In the given Transformer architecture for one Encoder Decoder block, identify every Add & Norm layer. Submit the corresponding component alphabets in alphabetical (ascending) order as a single uppercase string without spaces, commas, or quotation marks.

Example: If the correct components are Q, G, and S, your answer should be GQS

Correct answer: FHMPR

Source evidence: [{"page": 3, "bbox": [22.87, 335.0, 150.0, 365.0], "type": "explicit field", "detail": "Source section 'Possible Answers : FHMPR' for short answer alphanumeric response"}]

Question 3

Question ID: 6406532197544

Type: Multiple Choice Question · Marks: 1
PDF pages: 3

Shared context: ctx-2-4

Identify the function of component "L" in the given Transformer architecture.

Options

Positional Encoding

Source option ID: 6406537135075

Feed-Forward Network

Source option ID: 6406537135076

Masked Multi-Head Self-Attention

Source option ID: 6406537135077

Multi-Head Cross-Attention

Source option ID: 6406537135078

Linear Output Projection

Source option ID: 6406537135079

Correct answer: C

Source evidence: [{"page": 3, "bbox": [23.6200008392334, 532.1840209960938, 119.01998901367188, 548.5040283203125], "type": "green source option", "detail": "Source option 6406537135077 rendered in green font indicating correct option"}]

Question 4

Question ID: 6406532197545

Type: Multiple Select Question · Marks: 2
PDF pages: 3, 4

Shared context: ctx-2-4

Based on the architecture shown in the given image, identify all components that appear more than once. Assume the architecture contains exactly one Encoder block and one Decoder block.

Options

Positional Encoding

Source option ID: 6406537135080

Feed-Forward Network

Source option ID: 6406537135081

Masked Multi-Head Self-Attention

Source option ID: 6406537135082

Multi-Head Cross-Attention

Source option ID: 6406537135083

tokenizer

Source option ID: 6406537135084

Add & Norm

Source option ID: 6406537135085

Correct answer: A, B, F

Source evidence: [{"page": 3, "bbox": [23.6200008392334, 726.884033203125, 119.01998901367188, 743.2040405273438], "type": "green source option", "detail": "Source option 6406537135080 rendered in green font indicating correct option"}, {"page": 3, "bbox": [23.6200008392334, 747.7040405273438, 119.01998901367188, 764.0240478515625], "type": "green source option", "detail": "Source option 6406537135081 rendered in green font indicating correct option"}, {"page": 3, "bbox": [23.6200008392334, 38.324012756347656, 119.01998901367188, 54.64401626586914], "type": "green source option", "detail": "Source option 6406537135085 rendered in green font indicating correct option"}]

Question 5

Question ID: 6406532197546

Type: Multiple Select Question · Marks: 3
PDF pages: 4

A researcher is exploring an alternate approach to implement masked attention in the transformer decoder. The causal mask considered is a binary lower-triangular matrix $M$ (where $M_{ij} = 1$ for $j \le i$ and $0$ otherwise). This mask ($M$) is applied as an element-wise multiplication ($\odot$) after the softmax operation on Attention Matrix ($A$).

\[\text{softmax}(A) \odot M\]

Identify the correct statements from the following:

Question 5 prompt: masked attention implementation with element-wise causal mask M after softmax operation

Figure 2. Question 5 prompt: masked attention implementation with element-wise causal mask M after softmax operation

Options

The approach violates the autoregressive property

Source option ID: 6406537135086

The approach ensures autoregressive property is not violated by ensuring that
the final attention weights of future tokens are 0

Source option ID: 6406537135087

The final attention weights in the new approach will be greater than or equal
to that of the original approach

Source option ID: 6406537135088

The attention weights for the final token of the sequence will always be the
same for both the approaches

Source option ID: 6406537135089

Correct answer: A, D

Source evidence: [{"page": 4, "bbox": [23.6200008392334, 338.7039794921875, 119.01998901367188, 355.02398681640625], "type": "green source option", "detail": "Source option 6406537135086 rendered in green font indicating correct option"}, {"page": 4, "bbox": [23.6200008392334, 433.8039855957031, 119.01998901367188, 450.1239929199219], "type": "green source option", "detail": "Source option 6406537135089 rendered in green font indicating correct option"}]

Question 6

Question ID: 6406532197550

Type: Multiple Select Question · Marks: 3
PDF pages: 4

Which of the following options correctly reflect the internal state updates performed during a Byte Pair Encoding (BPE) training iteration?

Options

After a pair is chosen for merging, it is permanently added to the active
vocabulary list.

Source option ID: 6406537135102

The algorithm recalculates the log probability of the entire corpus before the
next iteration.

Source option ID: 6406537135103

The independent frequencies of the two constituent tokens that were part of
the merge are reduced by the frequency of the merged token.

Source option ID: 6406537135104

The original constituent tokens of the merged pair are removed from the
vocabulary to optimize dictionary size.

Source option ID: 6406537135105

Correct answer: A, C

Source evidence: [{"page": 4, "bbox": [23.6200008392334, 607.6840209960938, 119.01998901367188, 624.0040283203125], "type": "green source option", "detail": "Source option 6406537135102 rendered in green font indicating correct option"}, {"page": 4, "bbox": [23.6200008392334, 681.9640502929688, 119.01998901367188, 698.2840576171875], "type": "green source option", "detail": "Source option 6406537135104 rendered in green font indicating correct option"}]

Question 7

Question ID: 6406532197556

Type: Multiple Select Question · Marks: 3
PDF pages: 4, 5

A baseline T5 model with denoising objective on a span of corrupted tokens is used for unsupervised pre-training. Consider the phrase "life is like a box of chocolates". If the word "life" and the span "box of chocolates" are selected for corruption, which of the following statements correctly describe the training process?

Options

The input sequence replaces the corrupted spans with unique sentinal tokens
as: " is like a ".

Source option ID: 6406537135123

The output sequence generated takes the form " life box of chocolates
", with the final sentinal token to indicate sequence completion.

Source option ID: 6406537135124

The loss is calculated exclusively over the generated sentinal tokens and the
missing text.

Source option ID: 6406537135125

The decoder reconstructs the entire original sequence autoregressively to
compute the total loss over all positions.

Source option ID: 6406537135126

Correct answer: A, B, C

Source evidence: [{"page": 5, "bbox": [23.6200008392334, 124.42398834228516, 119.01998901367188, 140.74398803710938], "type": "green source option", "detail": "Source option 6406537135123 rendered in green font indicating correct option"}, {"page": 5, "bbox": [23.6200008392334, 161.5640106201172, 119.01998901367188, 177.88400268554688], "type": "green source option", "detail": "Source option 6406537135124 rendered in green font indicating correct option"}, {"page": 5, "bbox": [23.6200008392334, 198.7040252685547, 119.01998901367188, 215.02401733398438], "type": "green source option", "detail": "Source option 6406537135125 rendered in green font indicating correct option"}]

Question 8

Question ID: 6406532197547

Type: Multiple Choice Question · Marks: 3
PDF pages: 5

During the fine-tuning of GPT-1 for Textual Entailment tasks, what is the purpose of the delimiter token ($)?

Options

To trigger the softmax function.

Source option ID: 6406537135090

To separate premise and hypothesis.

Source option ID: 6406537135091

To mark the end of the sentence.

Source option ID: 6406537135092

To initialize the linear head.

Source option ID: 6406537135093

Correct answer: B

Source evidence: [{"page": 5, "bbox": [23.6200008392334, 488.5039978027344, 119.01998901367188, 504.8240051269531], "type": "green source option", "detail": "Source option 6406537135091 rendered in green font indicating correct option"}]

Question 9

Question ID: 6406532197548

Type: Multiple Choice Question · Marks: 3
PDF pages: 5, 6

During the fine-tuning of a decoder-only Transformer for sequence classification, why is the hidden representation of the final input token used as the input to the classification layer?

Options

The first token's representation relies exclusively on absolute positional
encodings and lacks semantic context for classification.

Source option ID: 6406537135094

The final token's representation is computed independently of self-attention,
making its gradients more stable during the fine-tuning process.

Source option ID: 6406537135095

Selecting the final token aligns with the autoregressive objective, eliminating
the need for teacher forcing during the classification forward pass.

Source option ID: 6406537135096

Causal self-attention ensures that only the final token's representation aggregates information from all preceding tokens in the sequence.

Source option ID: 6406537135097

Correct answer: D

Source evidence: [{"page": 6, "bbox": [23.6200008392334, 799.1240234375, 119.01998901367188, 815.4440307617188], "type": "green source option", "detail": "Source option 6406537135097 rendered in green font indicating correct option"}]

Question 10

Question ID: 6406532197549

Type: Multiple Choice Question · Marks: 3
PDF pages: 6

In the decoding phase of a Language model, temperature scaling with $T = 5$ is applied to the output logits before performing top-$p$ (nucleus) sampling. Which of the following options best describes the impact of the temperature scaling on the subsequent sampling step?

Question 10 prompt: temperature scaling parameter and top-p nucleus sampling effect on candidate token count

Figure 3. Question 10 prompt: temperature scaling parameter and top-p nucleus sampling effect on candidate token count

Options

Increases the required number of candidate tokens thus expanding the
nucleus.

Source option ID: 6406537135098

Decreases the required number of candidate tokens thus shrinking the
nucleus.

Source option ID: 6406537135099

Truncates the tail of the distribution causing the sampling to behave similar to top-$k$ sampling.

Source option ID: 6406537135100

Size of the nucleus remains unaffected by the temperature scaling as it only
depends on order of the tokens.

Source option ID: 6406537135101

Correct answer: A

Source evidence: [{"page": 6, "bbox": [23.6200008392334, 213.74400329589844, 119.01998901367188, 230.06399536132812], "type": "green source option", "detail": "Source option 6406537135098 rendered in green font indicating correct option"}]

Question 11

Question ID: 6406532197551

Type: Multiple Choice Question · Marks: 3
PDF pages: 6

In the WordPiece tokenization algorithm, how does the scoring formula prioritize token merges?

Options

It evaluates the log-probability of the pair and selects the highest scoring
transition.

Source option ID: 6406537135106

It merges the pair with the highest absolute frequency in the corpus.

Source option ID: 6406537135107

It prioritizes pairs that occur together frequently but rarely appear
independently in the corpus.

Source option ID: 6406537135108

None of these

Source option ID: 6406537135109

Correct answer: C

Source evidence: [{"page": 6, "bbox": [23.6200008392334, 568.1940307617188, 119.01998901367188, 584.5140380859375], "type": "green source option", "detail": "Source option 6406537135108 rendered in green font indicating correct option"}]

Question 12

Question ID: 6406532197555

Type: Multiple Choice Question · Marks: 3
PDF pages: 6, 7

Consider the T5 "text-to-text" paradigm. How does the model process fundamentally different tasks such as Semantic Similarity (continuous regression) and Sentiment Analysis (discrete classification) during the fine-tuning phase?

Options

The model uses the encoder to output the regression score and decoder for
the classification tokens switching based on the input prompt.

Source option ID: 6406537135119

The decoder part of the model is appended with separate prediction heads for
regression and classification which are updated based on Mean Squared Error and Cross-Entropy
respectively.

Source option ID: 6406537135120

The model treats both tasks as sequence generation objectives with task-
specific prefix added in the input and all output is generated as strings.

Source option ID: 6406537135121

None of these.

Source option ID: 6406537135122

Correct answer: C

Source evidence: [{"page": 7, "bbox": [23.6200008392334, 70.96402740478516, 119.01998901367188, 87.28402709960938], "type": "green source option", "detail": "Source option 6406537135121 rendered in green font indicating correct option"}]

Question 13

Question ID: 6406532197557

Type: Multiple Choice Question · Marks: 3
PDF pages: 7

In the context of the BART (Bidirectional AutoRegressive Transformer) architecture, how does the model handle corrupted input sequences compared to its output?

Options

The decoder ignores the encoder output and performs standard causal
language modeling.

Source option ID: 6406537135127

The encoder processes the original sequence and the decoder predicts the
corrupted tokens.

Source option ID: 6406537135128

The encoder processes the corrupted sequence and the decoder predicts the
entire original sequence.

Source option ID: 6406537135129

Both the encoder and decoder process the corrupted sequence to calculate
masked loss.

Source option ID: 6406537135130

Correct answer: C

Source evidence: [{"page": 7, "bbox": [23.6200008392334, 339.9440002441406, 119.01998901367188, 356.2640075683594], "type": "green source option", "detail": "Source option 6406537135129 rendered in green font indicating correct option"}]

Question 14

Question ID: 6406532197558

Type: Multiple Choice Question · Marks: 3
PDF pages: 7

When comparing the C4 dataset to an 'Unfiltered-C4' dataset that is 8 times larger, what was the observed effect on downstream task performance?

Options

Performance remained identical, suggesting a saturation point in data utility.

Source option ID: 6406537135131

Performance improved only for high-resource tasks like translation.

Source option ID: 6406537135132

Performance improved significantly due to the increased diversity of data.

Source option ID: 6406537135133

Performance degraded across all tasks despite the larger scale.

Source option ID: 6406537135134

Correct answer: D

Source evidence: [{"page": 7, "bbox": [23.6200008392334, 613.4240112304688, 119.01998901367188, 629.7440185546875], "type": "green source option", "detail": "Source option 6406537135134 rendered in green font indicating correct option"}]

Question 15

Question ID: 6406532197559

Type: Multiple Choice Question · Marks: 3
PDF pages: 7, 8

Which of the following describes the 'Adapter Layers' fine-tuning strategy?

Options

Randomly freezing 50% of the attention heads in each layer during training.

Source option ID: 6406537135135

Adding small dense-ReLU-dense blocks after FFN layers and updating only
those and Layer Norm parameters.

Source option ID: 6406537135136

Updating all parameters in the model but using a much smaller learning rate.

Source option ID: 6406537135137

Increasing the dimension $d_{\text{model}}$ of the existing Feed-Forward networks.

Question 15 Option D: increasing the dimension d_model of the existing Feed-Forward networks

Figure 4. Question 15 Option D: increasing the dimension d_model of the existing Feed-Forward networks

Source option ID: 6406537135138

Correct answer: B

Source evidence: [{"page": 7, "bbox": [23.6200008392334, 775.4840087890625, 119.01998901367188, 791.8040161132812], "type": "green source option", "detail": "Source option 6406537135136 rendered in green font indicating correct option"}]

Question 16

Question ID: 6406532197560

Type: Multiple Choice Question · Marks: 3
PDF pages: 8

Which of the following is a reported negative effect of including significantly duplicated content in LLM training data?

Options

It decreases the model's overall inference speed during deployment.

Source option ID: 6406537135139

It causes the model to generate repeated sequences much more frequently.

Source option ID: 6406537135140

It permanently reduces the maximum supported context window length.

Source option ID: 6406537135141

It causes the model to lose the ability to process multilingual input.

Source option ID: 6406537135142

Correct answer: B

Source evidence: [{"page": 8, "bbox": [23.6200008392334, 244.2040252685547, 119.01998901367188, 260.5240173339844], "type": "green source option", "detail": "Source option 6406537135140 rendered in green font indicating correct option"}]

Question 17

Question ID: 6406532197561

Type: Multiple Choice Question · Marks: 3
PDF pages: 8

If you have a fixed compute budget ($C$) and choose to train an extraordinarily large model (massive $N$) on a relatively small dataset (small $D$), what is the likely outcome according to scaling laws?

Question 17 prompt: scaling laws trade-off with fixed compute budget C, massive model size N, and small dataset D

Figure 5. Question 17 prompt: scaling laws trade-off with fixed compute budget C, massive model size N, and small dataset D

Options

The model will be over-trained and memorize the dataset.

Source option ID: 6406537135143

The model will be under-trained and less capable than a smaller model trained
on more data with the same compute budget.

Source option ID: 6406537135144

The model will perfectly memorize the small dataset and generalize flawlessly.

Source option ID: 6406537135145

The model's inference speed will increase dramatically.

Source option ID: 6406537135146

Correct answer: B

Source evidence: [{"page": 8, "bbox": [23.6200008392334, 502.9139709472656, 119.01998901367188, 519.2340087890625], "type": "green source option", "detail": "Source option 6406537135144 rendered in green font indicating correct option"}]

Shared context 18-19

Question ID: 6406532197552

Type: Comprehension
PDF pages: 8, 9

Consider the following corpus consisting of 4 words (ignore spaces and punctuation).

there is no spoon

The vocabulary is constructed using the whole words in this corpus and the individual characters in this corpus. The vocabulary should now have 13 tokens. Use this vocabulary for the given subquestions.

Question 18

Question ID: 6406532197553

Type: Multiple Select Question · Marks: 3
PDF pages: 9

Shared context: ctx-18-19

Using the vocabulary constructed, two new continuous strings: "prison" and "rhino" are to be segmented. Which of the following proposed segmentations are valid?

Options

prison

Source option ID: 6406537135110

p, r, is, o, n

Source option ID: 6406537135111

p, r, is, on

Source option ID: 6406537135112

r, h, i, n, o

Source option ID: 6406537135113

r, h, i, no

Source option ID: 6406537135114

Correct answer: B, D, E

Source evidence: [{"page": 9, "bbox": [23.6200008392334, 225.52403259277344, 119.01998901367188, 241.84402465820312], "type": "green source option", "detail": "Source option 6406537135111 rendered in green font indicating correct option"}, {"page": 9, "bbox": [23.6200008392334, 267.1639709472656, 119.01998901367188, 283.4839782714844], "type": "green source option", "detail": "Source option 6406537135113 rendered in green font indicating correct option"}, {"page": 9, "bbox": [23.6200008392334, 287.9839782714844, 119.01998901367188, 304.3039855957031], "type": "green source option", "detail": "Source option 6406537135114 rendered in green font indicating correct option"}]

Question 19

Question ID: 6406532197554

Type: Multiple Choice Question · Marks: 3
PDF pages: 9

Shared context: ctx-18-19

Using a unigram language model, identify the segment with the highest probability from the given options:

Options

n, o, i, s, e, s

Source option ID: 6406537135115

no, is, e, s

Source option ID: 6406537135116

nois, e, s

Source option ID: 6406537135117

no, i, s, e, s

Source option ID: 6406537135118

Correct answer: B

Source evidence: [{"page": 9, "bbox": [23.6200008392334, 461.8639831542969, 119.01998901367188, 478.1839904785156], "type": "green source option", "detail": "Source option 6406537135116 rendered in green font indicating correct option"}]

Answer key

Source question	Source identity	Correct answer	PDF pages
2	6406532197543	FHMPR	3
3	6406532197544	C	3
4	6406532197545	A, B, F	3, 4
5	6406532197546	A, D	4
6	6406532197550	A, C	4
7	6406532197556	A, B, C	4, 5
8	6406532197547	B	5
9	6406532197548	D	5, 6
10	6406532197549	A	6
11	6406532197551	C	6
12	6406532197555	C	6, 7
13	6406532197557	C	7
14	6406532197558	D	7
15	6406532197559	B	7, 8
16	6406532197560	B	8
17	6406532197561	B	8
18	6406532197553	B, D, E	9
19	6406532197554	B	9

Extraction notes

Status: verified
Count reconciliation: The portal section header declared 16 questions and 50 marks for Section 1 (LLM). The 16 declared items correspond exactly to the 16 top-level portal question units: Question 1 (administrative subject confirmation), Comprehension container (2 to 4), Questions 5, 6, 7 (Sub-Section 3), Questions 8, 9, 10, 11, 12, 13, 14, 15, 16, 17 (Sub-Section 4), and Comprehension container (18 to 19) (Sub-Section 5). Excluding Question 1 (confirmation) and unpacking the two comprehension containers yields exactly 18 top-level academic questions (Q2 to Q19). The 50 marks reconcile perfectly across all 18 academic questions: Sub-Section 2 [Q2 (2) + Q3 (1) + Q4 (2) = 5] + Sub-Section 3 [Q5 (3) + Q6 (3) + Q7 (3) = 9] + Sub-Section 4 [Q8 (3) + Q9 (3) + Q10 (3) + Q11 (3) + Q12 (3) + Q13 (3) + Q14 (3) + Q15 (3) + Q16 (3) + Q17 (3) = 30] + Sub-Section 5 [Q18 (3) + Q19 (3) = 6] = 5 + 9 + 30 + 6 = 50 marks.

Question 1 is the administrative subject confirmation, semantically excluded from academic question count while retaining full evidence.
Question 5 stem contains raster equation text in source PDF; high-resolution 288 DPI crop captured and formula transcribed as \text{softmax}(A) \odot M.
Comprehension 2 to 4 references an embedded Transformer encoder-decoder diagram; high-resolution crop captured as asset assets/ctx-2-4-passage-01-p002.png.
Questions 10, 15, and 17 contain inline mathematical symbols/parameters rendered as raster images in the source PDF; preserved high-resolution visual crops and restored mathematical notation in question text with explicit repairs.

Excluded source entries:

{"classification": "subject-confirmation", "page": 1, "bbox": [22.87, 735.0, 565.0, 842.0], "reason": "Administrative subject verification asking candidate to confirm registration and cross-check hall ticket for DEGREE LEVEL : LARGE LANGUAGE MODELS.", "identity": "6406532197541"}