Q1.What does full fine-tuning update, compared to PEFT methods like LoRA?
Q2.What does LoRA stand for?
Q3.What is the core idea behind LoRA?
Q4.What is "catastrophic forgetting" in the context of fine-tuning?
Q5.What is QLoRA primarily designed to reduce?
Q6.What is "instruction tuning"?
Q7.What does RLHF stand for?
Q8.What is the general purpose of RLHF / preference tuning?
Q9.What is DPO (Direct Preference Optimization) an alternative to?
Q10.When is fine-tuning generally NOT the right approach for adding new factual knowledge to a model?
Q11.What is a "PEFT" method, broadly?
Q12.Why does full fine-tuning require substantially more compute/memory than LoRA-based approaches?
Q13.What is a common minimum scale of data typically needed for a meaningful fine-tuning run?
Q14.What best characterizes prompting/in-context learning compared to fine-tuning?
Q15.What is a key risk of fine-tuning on a narrow, low-diversity dataset?
Q16.In LoRA, what happens to the original pretrained weights during training?
Q17.What is model distillation, and how does it relate to fine-tuning?
Q18.Why might a team choose PEFT/LoRA over full fine-tuning even when compute isn't a hard constraint?
Q19.What is the purpose of a held-out evaluation set when fine-tuning a model?
Q20.What does "adapter" typically refer to in the PEFT literature?
Q21.Why is learning rate a particularly sensitive hyperparameter when fine-tuning a pretrained LLM?
Q22.What is "prompt tuning," and how does it differ from full fine-tuning?
Q23.What's a common reason to prefer few-shot prompting over fine-tuning as a first attempt?
Q24.What does it mean for a fine-tuned model to be evaluated for "regression" against the base model?
Q25.Why is instruction-tuned data quality often considered more important than raw quantity?
