ML: Transformers

Question 1

Why were Transformers introduced?

Answer 1

Solve RNN limitations: sequential computation, vanishing gradients, long-range dependencies

Enable parallelization over sequence elements

Better scaling to large datasets

Question 2

How do Transformers achieve parallelization?

Answer 2

Self-attention computes relationships between all tokens simultaneously

Unlike RNNs, no sequential recurrence → GPU/TPU friendly

Batch computations are straightforward

Question 3

What are the main limitations of RNNs/LSTMs that Transformers overcome?

Answer 3

Sequential computation → slow

Difficulty with long-range dependencies

Vanishing/exploding gradients (residual connections, layer normalisation - stable activations, weight init)

Harder to scale to very large datasets

Question 4

Explain self-attention.

Answer 4

Computes attention scores between every token pair in a sequence

Weighted sum of values based on query-key similarity

Captures dependencies regardless of distance

Question 5

Why is positional encoding needed?

Answer 5

Self-attention is permutation invariant → needs explicit position info

Positional encodings add order information to input embeddings

Question 6

Types of positional encoding?

Answer 6

Sinusoidal (fixed) → allows extrapolation

Learned embeddings → trainable

Relative positional encodings → distance-aware attention

Question 7

What are the main components of a Transformer encoder?

Answer 7

Input embedding + positional encoding

Multi-head self-attention layer

Feedforward network

Residual connections + layer normalization

Question 8

What are the main components of a Transformer decoder?

Answer 8

Masked multi-head self-attention (prevents looking ahead)

Encoder-decoder attention

Feedforward network

Residual connections + layer normalization

Question 9

What is multi-head attention and why is it useful?

Answer 9

Multiple attention heads learn different relationships/subspaces

Concatenated and projected → richer representations

Helps model complex dependencies

Question 10

Difference between encoder-only, decoder-only, and encoder-decoder models?

Answer 10

Encoder-only (BERT): classification, embedding tasks

Decoder-only (GPT): autoregressive generation

Encoder-decoder (T5, BART): sequence-to-sequence tasks

Question 11

Why is layer normalization used instead of batch normalization?

Answer 11

Sequences have variable lengths → batch norm unstable

Layer norm normalizes per sequence element → stable training

Question 12

What are key training challenges for Transformers?

Answer 12

Quadratic attention complexity (memory/compute)

Need massive datasets

Overfitting if small data

Optimization instability → mitigated by Adam, learning rate warmup, gradient clipping

Question 13

How do modern Transformers improve training efficiency?

Answer 13

Sparse attention / efficient attention (Longformer, BigBird)

Mixed precision / float16

Gradient checkpointing

Model parallelism & pipeline parallelism

FlashAttention and memory-efficient kernels

Question 14

What is transfer learning in Transformers?

Answer 14

Pretrain on large corpora → fine-tune on downstream tasks

Often freezes lower layers to reduce computation and avoid overfitting

Question 15

What are the building blocks of LLMs?

Answer 15

Decoder-only Transformers (GPT, LLaMA) for autoregressive text

Large embedding matrices

Positional encodings

Layer normalization, feedforward networks, multi-head attention

Question 16

How do Transformers enable RAG (Retrieval-Augmented Generation)?

Answer 16

Query embedded → nearest document vectors retrieved

Retrieved context concatenated → fed to decoder

Decoder generates answer conditioned on retrieved information

Question 17

What problems do RAG pipelines solve?

Answer 17

Memory limitation of LLMs → can access external knowledge

Reduces hallucinations by grounding output in retrieved docs

Scalable knowledge updates without full model retraining

Question 18

How do embeddings interact with Transformers in agentic AI?

Answer 18

Embeddings used as queries/keys/values for retrieval

Clustered or indexed for efficient tool routing

Semantic similarity for planning / multi-step reasoning

Question 19

What are key improvements over vanilla Transformer?

Answer 19

Sparse / linear attention → Longformer, Performer

Relative positional encodings → improves long-range modeling

Adapter layers → efficient fine-tuning

Memory-augmented models → Retrieval-Enhanced Transformers

Question 20

How does attention scale with sequence length?

Answer 20

Standard self-attention: O(n^2 d) memory/time complexity

Sparse / local attention →
O(n⋅d) for long sequences

Question 21

What is instruction-tuning in LLMs?

Answer 21

Fine-tuning on human instructions → aligns model behavior

Often combined with RLHF (Reinforcement Learning from Human Feedback)

Improves agentic reasoning and task compliance

Question 22

How do Transformers handle multi-modal inputs?

Answer 22

Use modality-specific embeddings

Cross-attention layers combine text, images, audio

Examples: Flamingo, GPT-4 multi-modal, PaLI

Question 23

Sequence length is very long and self-attention uses too much memory. What do you do?

Answer 23

Use sparse attention or memory-efficient attention

Truncate or chunk sequences

Use Longformer / Performer / BigBird

Gradient checkpointing to save memory

Question 24

LLM generates incorrect facts. How does RAG help?

Answer 24

Retrieves up-to-date documents

Conditions generation on external knowledge

Reduces hallucination

Question 25

Training loss plateaus with larger model. Possible solutions?

Answer 25

Learning rate schedule / warmup Better initialization Larger batch size or gradient accumulation Mixed precision for efficiency Layer-wise learning rate decay

Question 26

You want the model to handle both text and images. What architecture changes?

Answer 26

Embed image patches (ViT-like) Cross-attention between modalities Multi-modal decoder or unified encoder-decoder Positional encodings for both modalities

Question 27

Attention scores cluster around 0.5 for all tokens. What could be wrong?

Answer 27

Queries/keys not properly scaled Input embeddings lack variation Softmax temperature issues Check layer normalization or initialization

Question 28

An agentic AI must plan multi-step tasks using LLMs. How do Transformers help?

Answer 28

Autoregressive decoding models multi-step reasoning Attention captures dependencies across steps Can integrate retrieved knowledge at each step (RAG) Memory modules track intermediate states

Question 29

What is fine-tuning in the context of LLMs?

Answer 29

Adjusting the weights of a pre-trained model on a smaller, task-specific dataset to specialize the model for a downstream task.

Question 30

Why is fine-tuning often preferred over training from scratch?

Answer 30

Reduces data and compute requirements Leverages pre-trained representations Faster convergence and better generalization Avoids overfitting when labeled data is small

Question 31

Name common strategies for fine-tuning large models.

Answer 31

Full model fine-tuning Adapter layers / LoRA (low-rank adaptation) Prefix-tuning / prompt-tuning Freeze early layers, train later layers only

Question 32

How do you evaluate fine-tuned models?

Answer 32

Task-specific metrics (accuracy, F1, BLEU, ROUGE, etc.) Human evaluation for subjective tasks Cross-validation or hold-out validation

Question 33

When is fine-tuning preferable to RAG?

Answer 33

Task is small and narrow domain High accuracy required on a specific dataset Model must produce output without external retrieval (low-latency scenarios)

Question 34

When is RAG preferable to fine-tuning?

Answer 34

Knowledge changes frequently Dataset is too large to fine-tune on Multi-domain tasks or open-domain QA Reduce compute/storage costs

Question 35

Can fine-tuning and RAG be combined?

Answer 35

Yes: Fine-tune model for alignment, instruction-following, or task adaptation Use RAG to provide up-to-date knowledge dynamically Often improves factuality and flexibility

Question 36

How does model size affect fine-tuning?

Answer 36

Large models → expensive, memory-heavy Parameter-efficient methods (LoRA, adapters) reduce cost

Question 37

How do you avoid catastrophic forgetting during fine-tuning?

Answer 37

Use small learning rates Freeze early layers Replay / mixed data Regularization (e.g., L2)

Question 38

What are key challenges of RAG pipelines?

Answer 38

Retriever quality → affects answer grounding Context length → long documents may exceed model input Latency in retrieval + generation Outdated or low-quality documents → hallucinations

Question 39

You need to update a chatbot daily with latest policies. Fine-tuning or RAG? Why?

Answer 39

RAG → updating document index is faster and cheaper than daily model retraining Fine-tuning daily is costly and impractical

Question 40

A legal model needs high precision on a small dataset. Fine-tuning or RAG?

Answer 40

Fine-tuning → small, domain-specific data, high accuracy needed RAG may introduce retrieval errors

Question 41

LLM provides outdated info. How do you fix it without retraining?

Answer 41

Use RAG → add/update documents in the retrieval database Model stays frozen; no retraining needed

Question 42

What is LoRA in the context of fine-tuning LLMs?

Answer 42

Low-Rank Adaptation (LoRA) is a parameter-efficient fine-tuning method where only low-rank updates to weight matrices are trained, while the original model weights remain frozen.

Question 43

How does LoRA reduce memory and compute cost?

Answer 43

Adds small low-rank matrices 𝐴,𝐵 to pre-trained weights 𝑊0 Only 𝐴,𝐵 are updated during training → fewer trainable parameters Avoids full model gradient computation and storage

Question 44

What types of layers are typically adapted with LoRA?

Answer 44

Linear projection layers (e.g., query/key/value projections in attention) Feedforward layers in MLP blocks of Transformers Layers most sensitive to task adaptation

Question 45

Why is LoRA effective for LLM fine-tuning?

Answer 45

Large models have redundant capacity → low-rank updates suffice for new tasks Preserves general pre-trained knowledge → reduces catastrophic forgetting Efficient storage → multiple task adapters can be stored separately

Question 46

Write the LoRA weight update formula.

Answer 46

W=W0​+ΔW,ΔW=AB

Question 47

How is rank 𝑟 chosen in LoRA?

Answer 47

Small 𝑟 → fewer parameters, faster training, may underfit Larger 𝑟 → more expressive, slower and more memory Tune empirically based on task complexity and data size

Question 48

How does LoRA compare to full fine-tuning in storage?

Answer 48

Full fine-tuning: stores full model per task (~billions of params) LoRA: stores only low-rank adapters (~few million params) Multiple adapters can share a single frozen backbone

Question 49

Can LoRA be combined with other parameter-efficient methods?

Answer 49

Yes, e.g.: LoRA + prompt-tuning LoRA + adapters LoRA + prefix-tuning Useful for multi-task learning and modular AI agents

Question 50

When is LoRA particularly useful?

Answer 50

Large models (GPT, LLaMA, Falcon) where full fine-tuning is expensive Multiple task-specific adaptations Situations with limited compute or memory

Question 51

How do you evaluate LoRA adapters?

Answer 51

Standard downstream task metrics (accuracy, F1, BLEU, ROUGE) Compare to baseline pre-trained model without adapters Optionally compare to full fine-tuning

Question 52

Can LoRA adapters be swapped or shared across tasks?

Answer 52

Yes, because base weights remain frozen Adapters are modular → can dynamically switch tasks without retraining backbone

Question 53

You fine-tune an LLM on domain-specific QA using LoRA, but it underperforms. What can you do?

Answer 53

Increase LoRA rank 𝑟 Train more epochs or increase learning rate Check if the adapted layers are appropriate (attention vs feedforward) Combine with instruction-tuning or RAG for additional context

Question 54

ou want to support multiple domains on one LLM efficiently. How does LoRA help?

Answer 54

Store one frozen backbone Train small adapters for each domain Load the corresponding adapter at inference → low storage and compute

Question 55

Why is masking needed

Answer 55

Padding masks - for handling variable lengths Computers love consistency, but human sentences vary in length. When we process sentences in batches, we pad shorter sentences with zeros (or a token) to make them the same length as the longest one.The Problem: The attention mechanism doesn't know these pads are "garbage" data. It would try to calculate relationships between actual words and empty padding tokens.The Solution: A Padding Mask tells the model to ignore those specific positions by setting their attention scores to $-\infty$ before the Softmax layer. This ensures the padding has zero influence on the final output. Causal (look-ahead) masks - preventing cheating When a model is predicting the next word in a sentence, it shouldn't be allowed to see the words that come after the current position.The Problem: During training, we give the model the entire finished sentence. If we didn't use a mask, the model would simply "look ahead" at the next word in the input and copy it, rather than learning how to predict it.The Solution: We apply a triangular mask (often called a Causal Mask). For any word at position $i$, the mask blocks out all words at positions $i+1, i+2, \dots, n$ Mathematically, apply +M before softmax.