(editing) Data Science Interview Prep - LLM/AI

📚 The Start of Recording TIL Summer ‘24 : Why Did I Start This Project? 🚀

🔔 Don't Miss: Continous Records of TIL '25 (Still On-Going!) 🌟

✨ A New Chapter in TIL '25: Reflecting, Growing, and Moving Forward 📖

Interview Questions & Answers

LLM & AI

---

Questions from:

LinkedIn Post by Daniel Lee : LLM Interview Questions

---

LLM

Basic Concepts - LLM

• Architecture & Training
  • Transformer Architecture (attention mechanisms
  • Pre-training vs Fine-tuning
  • Training Objectives (next token prediction)
  • Context Window and Position Embeddings
  • Tokenization Strategies
  • Model Scaling Laws
  • Parameter Efficient Fine-tuning (LoRA, QLoRA, Prefix Tuning)
• Generation & Control
  • Temperature and Top-p Sampling
  • Prompt Engineering Techniques
  • Few-shot Learning
  • In-context Learning
  • Chain-of-Thought Prompting
  • Hallucination Prevention

---

Answers

Architecture & Training

1. Transformer Architecture (Attention Mechanisms):
   • Transformers use self-attention to weigh the importance of each word in a sequence, allowing the model to capture dependencies regardless of distance. This architecture has led to breakthroughs in NLP by effectively modeling long-range context.
2. Pre-training vs. Fine-tuning:
   • Pre-training involves training on vast datasets to develop general language understanding. Fine-tuning adapts the model to specific tasks with a smaller, domain-specific dataset, enhancing performance in target areas.
3. Training Objectives (Next Token Prediction):
   • For language models, their main goal is to predict the next token in a sequence. This capability fosters the learning of contextual relationships within the text, enabling the model to produce sequences that are both **coherent** and **accurate** in context.
4. Context Window and Position Embeddings:
   • The context window defines the maximum token length the model can process at once. Position embeddings retain token order because Transformers inherently do not depend on the order of sequences.
5. Tokenization Strategies:
   • Tokenization splits text into smaller units (tokens) for model processing. Common strategies include Byte Pair Encoding (BPE) and WordPiece, which balance vocabulary size and generalization.
   • Byte Pair Encoding (BPE) is a tokenization method that compresses text by iteratively merging the most frequent pairs of characters or subwords. This approach creates a manageable vocabulary of subword units, allowing models to handle rare or compound words by breaking them into smaller, reusable tokens.
6. Model Scaling Laws:
   • Scaling laws reveal that larger models (with more parameters and data) typically perform better, but with diminishing returns. These laws guide resource allocation for optimal model size and training.
7. Parameter Efficient Fine-tuning (LoRA, QLoRA, Prefix Tuning):
   • Techniques like LoRA, QLoRA, and Prefix Tuning enable fine-tuning large models efficiently by adjusting only specific layers or adding lightweight adapters, making the process faster and more cost-effective.
   • LoRA (Low-Rank Adaptation): LoRA fine-tunes large models by injecting low-rank matrices into specific layers, adjusting only a few parameters. This makes the process more efficient by minimizing memory and computational requirements.
   • QLoRA (Quantized LoRA): QLoRA combines low-rank adaptation with quantization, reducing model precision to save memory and computing costs. It’s beneficial for deploying large models on resource-constrained devices.
   • Prefix Tuning: Prefix Tuning fine-tunes models by adding a learned "prefix" of virtual tokens to the input, which guides the model in task-specific ways without modifying the core model parameters. This method is lightweight and effective for many NLP tasks.

---

Generation Control

1. Temperature and Top-p Sampling:
   • Temperature controls randomness in generation (lower values make the output more focused, and higher values increase creativity). Top-p sampling (nucleus sampling) selects tokens from the smallest set whose cumulative probability meets a threshold, reducing unlikely outcomes.
2. Prompt Engineering Techniques:
   • Crafting prompts to guide model responses improves task-specific performance. Good prompts often include explicit instructions, context, or examples to direct the model effectively.
3. Few-shot Learning:
   • Few-shot learning uses a handful of examples within a prompt to enable the model to generalize to similar tasks without extensive re-training, leveraging prior knowledge from pre-training.
4. In-context Learning:
   • In-context learning involves showing the model a context or sequence that influences its response, allowing it to generate relevant output based on preceding examples or information.
5. Chain-of-Thought Prompting:
   • This method encourages the model to think step-by-step, improving reasoning in complex tasks by breaking down its response into logical, sequential steps, which enhances interpretability and accuracy.
6. Hallucination Prevention:
   • Techniques like fact-checking, retrieval augmentation, and reinforcement tuning reduce hallucinations, ensuring the model’s responses are accurate and grounded in factual information.