Day 14 Generative AI Quiz: Test Your LLM Knowledge

Correct Answer: B
Explanation:
Sliding window attention (e.g., in Longformer) limits each token's attention to a fixed window of adjacent tokens. This reduces memory from O(n²) to O(n*k) (k=window size), making it ideal for long sequences.

Best Practice:

• Use sliding windows for document-level tasks
• Combine with global attention for key tokens (e.g., [CLS] token)
• Example: LongformerModel in Hugging Face uses 4,096-token windows

Common Mistake: Full attention (A) exhausts memory for long sequences (>512 tokens).

Correct Answer: B
Explanation:
Temperature (τ) controls randomness in softmax: softmax(logits / τ). Higher τ (>1.0) flattens probabilities for diverse outputs, while lower τ (<1.0) sharpens distributions.

Best Practice:

• Start with τ=0.7 for balanced creativity
• For technical content: τ=0.2–0.5
• For creative writing: τ=0.8–1.2
• Code example:

  generate(input_ids, temperature=0.8, do_sample=True)

Correct Answer: B
Explanation:
Adversarial debiasing trains the model to make predictions while minimizing an adversary's ability to detect protected attributes (e.g., gender).

Best Practices:

1. Pretrain with balanced datasets (e.g., Wikipedia + counterfactual augmentation)
2. Use loss terms like:

   loss = task_loss - λ * adversary_loss

3. Test with benchmarks: BOLD, HONEST

Common Pitfall: Prompt engineering (C) masks symptoms but doesn't fix model weights.

Correct Answer: D
Explanation:
LoRA freezes pretrained weights and injects trainable low-rank matrices into attention layers. It achieves ~95% of full fine-tuning accuracy with 10x fewer parameters.

Best Practice:

• Use r=8 (rank) for most tasks
• Target only attention matrices:

  peft_config = LoraConfig(r=8, target_modules=["q_proj", "v_proj"])

• Example: peft.LoraModel library

Why Not (C)? Full fine-tuning risks catastrophic forgetting with small datasets.

Correct Answer: A
Explanation:
FlashAttention uses tiling and recomputation to reduce memory reads/writes and significantly speeds up computation for long sequences. It’s 3x faster for sequences >2K tokens.

Best Practice:

• Enable via model(..., use_flash_attention=True)
• Combine with ALiBi (D) for extrapolation beyond trained context
• Hardware requirement: CUDA kernels

Real-World Use: GPT-4 uses FlashAttention-2 for 32K context.

Correct Answer: D
Explanation:
Reward models inherit biases from human preference data. If annotators favor lengthy responses, the model learns to prioritize verbosity over correctness.

Best Practices:

• Curate preference pairs with unambiguous criteria
• Use ranking systems (e.g., Elo) instead of absolute scores
• Annotator training example:

  "Rate accuracy over style; penalize factual errors"

Solution: Augment data with correctness-focused samples.

Correct Answer: C
Explanation:
Unigram tokenization (e.g., SentencePiece) probabilistically drops subwords, better handling rare/morphologically rich words.

Best Practice:

• Vocabulary size: 100K–200K for multilingual models
• Example for Finnish:

  tokenizer = SentencePieceProcessor(model_file="spm_finnish.model")

• Avoid BPE (A) for agglutinative languages—it fragments words excessively.

Correct Answer: C
Explanation:
Top-p (nucleus) sampling dynamically selects the smallest set of tokens whose cumulative probability ≥ p. Using p=0.95 includes plausible alternatives while excluding low-probability nonsense.

Best Practice:

• Top-p > 0.9 for creative tasks
• Disable top-k (k=0) to rely solely on probability mass
• Avoid low top-p values (A,D): overly restrictive
• Configuration:

  generate(inputs, top_p=0.95, top_k=0, temperature=0.8)