Skip to content

Fine-Tuning & Adaptation interview questions

PEFT, LoRA, QLoRA, RLHF, DPO, GRPO, quantization (GGUF, AWQ, GPTQ), model distillation, and domain adaptation.

What to study

  • LoRA rank r and alpha scaling factor tuning.
  • Preference optimization: RLHF vs DPO vs GRPO.
  • Quantization techniques: INT8/INT4 weight-only vs KV cache quantization.

Questions and answers

  1. What is LoRA and QLoRA for efficient fine-tuning?

    Intermediate · 2 min read

  2. What is the difference between pre-training, fine-tuning, and in-context learning?

    Beginner · 2 min read

  3. What is quantization and how do INT8 and INT4 reduce LLM footprint?

    Beginner · 2 min read

  4. How does Direct Preference Optimization (DPO) differ from RLHF?

    Intermediate · 2 min read

  5. How does Group Relative Policy Optimization (GRPO) work in DeepSeek R1?

    Advanced · 2 min read

  6. What is Supervised Fine-Tuning (SFT) and when is it required?

    Beginner · 2 min read

  7. What is catastrophic forgetting during LLM fine-tuning?

    Beginner · 2 min read

  8. What is LoRA rank r and alpha scaling factor?

    Beginner · 2 min read

  9. How does model distillation transfer knowledge from teacher to student?

    Intermediate · 2 min read

  10. How does Kahneman-Tversky Optimization (KTO) differ from DPO?

    Advanced · 2 min read

  11. What is dataset formatting for instruction tuning (Alpaca vs ShareGPT formats)?

    Beginner · 2 min read

  12. What is learning rate scheduling (cosine decay) during LLM fine-tuning?

    Beginner · 2 min read

  13. What is gradient accumulation and how does it simulate larger batch sizes on small GPUs?

    Beginner · 2 min read

  14. What is mixed precision training (FP16 vs BF16) and why is BF16 preferred on modern GPUs?

    Beginner · 2 min read

  15. What is weight merging in LoRA and why does it eliminate inference latency penalties?

    Beginner · 2 min read

  16. How does DoRA (Weight-Decomposed Low-Rank Adaptation) improve directional weight updates over LoRA?

    Intermediate · 2 min read

  17. How does GGUF format enable quantization and CPU/GPU offloading in llama.cpp?

    Intermediate · 2 min read

  18. How does AWQ (Activation-aware Weight Quantization) preserve critical weights compared to GPTQ?

    Intermediate · 2 min read

  19. How does ORPO perform SFT and alignment in a single step without reference models?

    Advanced · 2 min read

  20. How does DeepSpeed ZeRO stage 1, 2, and 3 partition optimizer states, gradients, and parameters?

    Advanced · 2 min read

Study Fine-Tuning & Adaptation with progress tracking