Question 1
Only ''apple''
"apple'' or ''banana''
"apple'', ''banana'' or ''grape''
"apple''or ''grape''
Large Language Models End Term: 21 December 2025, Set 1-2 (September 2025 term)
The IIT Madras BS Large Language Models (LLM) End Term paper sat on 21 Dec 2025, in the September 2025 term, set 1-2: 25 questions for 50 marks in 180 minutes. Every question is below with its answer. Take it as a timed mock test to be marked, or read it through first.
- 25
- 50
- 180 min
- 17
- 4
- 4
Show answer
Correct answer
"apple'' or ''banana''
Question 2
Both add 0 trainable parameters
Show answer
Correct answer
Question 3
Insufficient information to determine
Show answer
Correct answer
Insufficient information to determine
Question 4
For a 4-category classification task, how would you use BERT's output?
Use all token embeddings and average them.
Use the [CLS] token embedding as input to a classification layer.
Use only the [SEP] token
Use the [MASK] token predictions.
Show answer
Correct answer
Use the [CLS] token embedding as input to a classification layer.
Question 5
A hospital wants to extract disease mentions and treatment plans from unstructured clinical notes. Notes use heavy abbreviations like ``pt dx w/ CAD, rx aspirin'' For extracting specific disease names from text, which task should you fine-tune BERT for?
Masked Language Modeling
Next Sentence Prediction
Named Entity Recognition
Text generation
Show answer
Correct answer
Named Entity Recognition
Question 6
Show answer
Correct answer
Question 7
Show answer
Correct answer
Question 8
Show answer
Correct answer
Question 9
If a model uses RoPE, increasing the sequence length from 100 to 1000 requires:
900 more positional vectors
No new parameters
A new embedding matrix
Show answer
Correct answer
No new parameters
Question 10
Show answer
Correct answer
Question 11
APE = 6, RPE = 11
APE = 6, RPE = 6
APE = 12, RPE = 11
APE = 5, RPE = 12
Show answer
Correct answer
APE = 6, RPE = 11
Question 12
A model with positional encoding crashes when the input length exceeds the training max length. Which encoding is it most likely using?
ALiBi
APE
RoPE
RPE
Show answer
Correct answer
APE
Question 13
Show answer
Correct answer
Question 14
Which of the following decoding method(s) can generate a different output every time you run the model with the same prompt?
Greedy decoding
Beam search with beam width = 1
Top-p sampling with temperature T = 1
Show answer
Correct answers
Top-p sampling with temperature T = 1
Question 15
RoPE
APE
ALiBi
None of these
Show answer
Correct answer
ALiBi
Question 16
Show answer
Correct answers
Question 17
Show answer
Correct answer: 32
Question 18
Which of the following statements accurately describes the properties and computational complexities of the various sparse and block attention mechanisms?
Dilated Attention increases the computational complexity significantly compared to standard Strided Local Attention because it covers a wider receptive field.
Global, Random, and Local Window attention patterns handle long sequences well.
Empirical observations suggest that random permutations of blocks are significantly more important for performance than the identity permutation.
Show answer
Correct answers
Global, Random, and Local Window attention patterns handle long sequences well.
Question 19
Show answer
Correct answer: 144
Question 20
Suppose you are working on prefix language modeling. The sequence length is 10 and the first two tokens represent the task-specific prefix. How many non-infinity elements are there in the mask for computing attention scores?
Show answer
Correct answer: 56
Question 21
Show answer
Correct answer: 262144
Question 22
During training, they strictly enforce teacher forcing, meaning the ground-truth previous tokens are always provided as input to predict the current token. Based on the above setup, answer the given subquestions.
Which of the following describes the fundamental training mechanism of this model?
Show answer
Correct answer
Question 23
During training, they strictly enforce teacher forcing, meaning the ground-truth previous tokens are always provided as input to predict the current token. Based on the above setup, answer the given subquestions.
It dictates that the loss is only calculated for tokens that were incorrectly predicted.
Show answer
Correct answer
Question 24
During training, they strictly enforce teacher forcing, meaning the ground-truth previous tokens are always provided as input to predict the current token. Based on the above setup, answer the given subquestions.
In the inference phase (when the model is used for generation), why is the teacher forcing technique not used, and what is the primary challenge this introduces?
Teacher forcing is used, but only with a small probability.
It's not used because the ground-truth sequence is unknown; this introduces the exposure bias problem, where the model is unfamiliar with its own prediction errors.
It's not used because it drastically slows down the inference speed.
It's not used; the challenge is the need for a separate encoder-decoder architecture.
Show answer
Correct answer
It's not used because the ground-truth sequence is unknown; this introduces the exposure bias problem, where the model is unfamiliar with its own prediction errors.
Question 25
During training, they strictly enforce teacher forcing, meaning the ground-truth previous tokens are always provided as input to predict the current token. Based on the above setup, answer the given subquestions.
Which component of the Transformer decoder is essential for enforcing the Causal Language Modeling constraint?
The Feed-Forward Network layer.
The Positional Encoding layer.
The masked Self-Attention layer.
The Cross-Attention layer.
Show answer
Correct answer
The masked Self-Attention layer.